Abstract

Artificial intelligence includes the subfield of deep learning. Recent years have seen
increased attention from the academic and business communities of autonomous learning and
feature extraction. It has found a lot of usage in areas like NLP, speech recognition, and video
analysis. At the same time, it has become a centre of attention for scientists studying methods
of agricultural plant protection, such pest range assessments and the study of plant diseases.
Researchers may avoid the pitfalls of selecting disease spot qualities subjectively, derive
more accurate plant disease features, and speed up the rate of technological transformation all
by disease detection in plants using deep learning. In this study, we will take a look at how
far deep learning technology has come in the field of identifying diseases in crop leaves
during the past few years. In this work, we discuss the cutting edge of diagnostics for leaf
diseases utilizing both deep learning and cutting-edge image methods, as well as the obstacles
that must be overcome.

1. Introduction

As one of the world's largest sectors, agriculture provides vital support in the forms of
food, income, and employment. Since there are so many farmers in India and other low- and
middle-income nations, agriculture accounts for 18% of national income and 53% of total
employment [1]. Over the previous three years, agriculture's share of the country's GVA has
risen from 17.6% to 20.2% [2,3]. The majority of the expansion of the economy may be
attributed to this field. Therefore, the quality of food production could be negatively impacted
if plant diseases and insect infections had an effect on agriculture. The rising number of
people need more nourishment. Increasing agricultural output and ensuring crop security are
essential to meeting this urgent demand. However, a wide variety of pathogens in crops'
natural habitats makes them vulnerable to a wide range of diseases. Disease-causing
microorganisms include viruses, fungi, bacteria, and others [1]. The quantity and quality of
food produced might suffer greatly when crop diseases lower output by as much as 95% [2].
In order to prevent massive losses and cut down on the overuse of potentially dangerous
pesticides, early disease detection is essential. Small farmers and those in underdeveloped
nations sometimes rely solely on visual symptoms to diagnose crop illnesses. This is a
laborious job that calls for knowledge of plant pathology and extensive time for treatment [3].
Additionally, if a rare disease is attacking the field, farmers will seek expert help to get an
accurate and effective diagnosis, which will certainly lead to higher treatment expenses [4].
Therefore, huge farms can't rely on visual inspection, and it can even lead to inaccurate
projections because of human error [5]. As consumer demand rises, businesses must find
ways to reduce the harmful effects of chemical intake on the environment and human health,
researchers have developed technological recommendations for early detection in crops in an
accurate, quick, and secure manner.

Several approaches [6,7,8,9] have been put forth with the aim of automating the detection
of diseases. Both direct and indirect approaches have been developed for automatic disease
detection in crops [10]. Direct methods include molecular [11] and serological [12,13]
techniques, which allow for precise and direct detection of the pathogens causing the disease,
but which take a long time to collect, process, and analyze the samples. In contrast, indirect
methods, like as optical imaging techniques [14,15], can detect illnesses and forecast crop
health by analyzing variables including changes in morphology and rate of breathing out.
Two of the most common indirect techniques for early illness detection include fluorescence
and hyperspectral imaging [16]. However, low-income farmers sometimes have a hard time
obtaining hyperspectral devices due to their high cost, cumbersome size, and limited
availability, despite the fact that hyperspectral photographs are a vital source of information,
because they provide more details than standard images do[17]. There are many other digital
cameras available at reasonable costs in retailers like Best Buy and Amazon. So far, most
automatic identification methods have handled images in the visible domain, where very
precise and quick techniques can be applied.

Visual crop inspection can be a useful tool for assessing quality, but it can be time-
consuming, expensive, and imprecise. Researchers have developed a number of methods,
such as those used in object recognition and image analysis for quality control purposes. In
order to identify and categorize agricultural diseases [1-3], this paper use image processing
technologies. High-resolution images, which are notoriously difficult to get, are essential for
this image processing method's identification and categorization of illnesses. Because of this,
disease prediction is not only difficult, but also time-consuming. The most commonly used of
conventional crop disease detection methods, manual observation, has been shown to be
ineffective and unreliable. Due to a lack of professional expertise, farmers and the limited
availability of agricultural professionals, they often fail to take preventative measures until it
is too late. In recent years, advancements have been made in various fields, including image
processing [1], pattern recognition [2], computer vision [3], and others. Using computers to
automatically identify diseases is a useful step in solving agricultural problems.
 Image pre-processing, researcher-designed complicated disease characteristics for feature
extraction [5], and machine learning algorithms [6] for identifying crop illnesses constitute
the three pillars of the conventional machine vision [4] approach to detecting diseases in crop
leaves. Many different supervised machine learning classification methods, including NB [5],
KNN [6, 7], DT [8], SVM [9], and RF [10, 11], are discussed for identifying and classifying
diseases from plant leaves, and a comparison is done between them. Additionally, it offers a
variety of methods that, when compared to others, produce the most precise results. Several
domains [12] benefit from the use of these classification strategies, including biological
signal processing [13] and healthcare [14, 15].

Improvements in plant disease recognition research have been made thanks to the use of
deep learning technologies in recent years. Since deep learning (DL) technology is hidden
from the end user, researchers in plant protection and statistics don't need to have a high level
of expertise to use it. With DL, researchers recognizes and labels plant disease areas based on
photo characteristics, saving time and effort over manual processes. For these reasons, the
application of deep learning to the problem of plant disease recognition has recently been a
focal point of scientific inquiry. Larger datasets, flexible multicore GPUs, and progress in
training deep neural networks and ancillary software libraries like NVIDIA's computing
unified device architecture (CUDA) are all contributing reasons.

2. Crop disease and their symptoms

This section discusses the numerous methods used to diagnose plant illnesses, lays out a
taxonomy of those diseases, and explains what image processing and machine learning are. It
also shows how hyperspectral images, the IoT, deep learning, and transfer learning may be
used for disease diagnosis. The disease's symptoms, which include stunted crop growth, are
readily apparent. The first sign of disease in plants is a change in leaf colour. In addition, the
structure and feel of the leaves can tell a lot about the presence and severity of a disease.
Thus, leaf photos can be processed to detect illnesses like mildew, rust, and powdery mildew
[26,27].
 The three categories of plant diseases [28] seen in Figure 1 and defined in Table 1 below
are briefly discussed below:
 Virus diseases [1]: Infectious plant illnesses are particularly challenging to diagnose since
there is no reliable signal that can be used to track their progression over time. As a result,
their symptoms are often misinterpreted as indications of nutritional inadequacy or injury.
Cucumber crawling bugs, aphids, whiteflies, and leafhoppers are all vectors for virus
transmission.
 Fungal diseases [1]: Infectious fungi are to blame for a wide variety of foliar diseases,
including downy mildew, anthracnose, and powdery mildew. Old, lower leaves that are
spotted with grayish green or are drenched in water are the first to show symptoms. These
areas get darker and develop fungus as the parasite develops.
 Bacterial diseases [1]: Vegetables are extremely susceptible to pathogen-caused illnesses.
They gain access to the plant life indirectly, through cracks or holes. Insects, diseases, and
even the tools used to perform routine agricultural chores like harvesting and trimming
can cause damage to crops.
3. Research Questions
The purpose of this paper is to identify the machine learning approaches currently used
for disease detection in crops. In addition, it seeks to identify the most recent and cutting-
edge machine learning methods that have been implemented. This leads us to our primary
study topic, which is:
PRQ: “What are the state-of-the-art machine learning approaches utilized for the problem
of detecting crop diseases during recent years?”
To assist in refining the focus of the original research topic, a set of secondary questions
is also provided:
SRQ2: Which crop diseases cause the most damage and are most common?
SRQ3: What kinds of data sets are there to choose from?

SRQ4: How do crop disease detection experts often measure success?

 SRQ5: What are the most popular machine learning frameworks?

4. Machine and Deep Learning for Plant disease classification

In the area of image processing, convolutional neural networks have demonstrated

remarkable success as feature classifiers and extractors during the past decade. While this
idea has been effectively implemented elsewhere, it is just recently making its way into the
agricultural sector. This section provides a high-level summary of the various applications of
machine learning in plant disease categorization, including the identification of weeds and
weed infestations, the counting of flowers and fruits, and the grading of harvested produce.
Since 2015, DL has seen extensive use in the field of diagnosing leaf diseases. DL
representation strategy that, rather than focusing on representing semantic aspects, tries to
represent data optimization approaches optimally. Features are extracted automatically as
opposed to manually as a result of the learning process. Also, DL is a part of the cutting-edge
farming practices that will help the food business develop in areas like automation,
advertising, massive amounts of data, pest control, healthcare diagnosis, and technology.

Ramesh et al. [1] produced data sets for use with Random Forest to determine which leaves
were healthy and which were infected. In order to categorize photos of diseased and healthy
leaves, a random forest classifier is trained using data sets of both types of leaves. They used
the Histogram of an Oriented Gradient (HOG) to pick out features of images.

Harakannanavar et al. [2] assessed the affected tomato leaves samples. These tomato leaves
disorder samples will help farmers detect diseases early and protect their crops. The correct
diagnosis and categorization of these illnesses is essential. Tomato harvests can be impacted
by a number of different illnesses. If these illnesses could be detected sooner, they could have
less of an impact on tomato plants, leading to a higher crop output. Many novel approaches
have been explored to detect and categorize various diseases. The goal of this endeavour is to
aid farmers in making correct diagnoses of diseases in their earliest stages. To successfully
characterize and categorize tomato illnesses, the CNN is utilized. The input photos are first
subjected to preprocessing, during which the desired regions are extracted from the full-
resolution versions. Second, the photos are processed further by playing around with different
CNN model hyper-parameters. Finally, CNN is able to extract additional features from
images, such as color, texture, edge, and more.
Chug et al. [3] presented a revolutionary framework that merges the benefits of machine
learning and deep learning. Forty Hybrid Deep Learning models are included in the suggested
framework. These models are a combination of eight built-in deep learning framework
variants (EfficientNet B0-B7) used as feature extractors.

A support vector machine (SVM) [84] is a model for generating a hyper-plane that divides a
problem space into many classes. SVM selects the best hyper-plane to separate the data by
increasing the separation between the most closely spaced data points of every category to the
hyper-plane. Using kernel trick technique, SVM may also perform well on non-linear data.
The SVM kernel is a function that maps input data from a low-dimensional space to a linearly
separable space in a higher dimension. This is why support vector machines (SVMs) can
perform so well in high dimensional spaces. Regression issues are another common
application of SVM [29,40,85]. In addition, Bhatia et al. [86] demonstrated a hybrid use of
SVM by combining it with a logistic regression approach to forecast the occurrence of
powdery mildew on tomato plants. Table 3 provides a summary of research conducted in the
agricultural sector using SVM as the ML model. It is possible to see the SVM type, kernel,
and output. Most applications of SVM-based classification and regression algorithms in
agricultural settings make use of linear, polynomial, or RBF kernels.
Saleem et al. [4] implemented the NZDLPlantDisease-v1 dataset. To identify plant disease
using the recently generated dataset, an improved version of the best attained deep learning
(DL) model, region-based fully convolutional network (RFCN), has been developed. The
data augmentation methods were assessed one by one once the optimal DL model was
determined. After that, they looked into the results of DL optimizers, batch normalizers,
weight initializers, and image resizers using interpolators. After that, position-sensitive score
corresponds and anchor box parameters were observed empirically to improve performance.
Additionally, a layered k-fold cross-validation method and testing on an external dataset were
used to show the practicality of the suggested approach.

Nandhini et al. [5] applied CNN and RNN to agricultural disease categorization and
detection, adding to their already impressive track record of success in other areas. Building a
Deep Learning Model for disease categorization and early prediction to aid plantain tree
farmers is the focus of their research. By fusing RNN with CNN, the Gated-Recurrent
Convolutional Neural Network (G-RecConNN) is put forward as a new progressive image
classification model for illness detection. Plant picture sequences are used as input in the
suggested model.

Arun et al. [6] provided a multi-crop disease detection model that can categorize crop
diseases across crop types using the Complete Concatenated Deep Learning (CCDL)
architecture. Complete Concatenated Blocks (CCBs) are introduced as fundamental building
blocks in this design. To limit the total number of model parameters, the point convolution
layer is placed at the beginning of each convolution layer in this component. The CCB's
convolution layers are subjected to a full concatenation path. It improves feature map use and
hence classification accuracy. The restructured Plant Village dataset is used to train the
suggested architecture. Pruned Complete Concatenated Deep Learning (PCCDL) models are
trained models that have been reduced in size through pruning.

Chen et al. [7] improved the model's ability to detect subtle plant lesion traits by discussing
deep learning techniques and developing a convolutional mixed network. To better plant
disease recognition, they combined three lightweight convolutional neural networks (CNNs)
using ensemble learning to create a new network they named Es-MbNet. Model training
opted for a two-stage approach based on transfer learning, with focusing on the setup of
network weights. Using the weights learned in the first phase of training, the network was
retrained in the second using the target dataset to obtain the best possible model parameters.
Hassan et al. [8] suggested a variety of deep learning models for use in diagnosing plant
diseases. However, deep learning models call for a lot of parameters, making the training
period longer and the implementation tricky on mobile devices. A new deep learning model
utilizing the inception layer and residual connection has been suggested. One method for
doing this is by employing depth-separable convolution.

Thakur et al. [9] developed the 'VGG-ICNN', a lightweight CNN designed to detect crop
illnesses from photographs of plant leaves. Compared to other high-performing deep learning
models, VGG-ICNN has a lot less parameters (about 6 million) to work with.

Abd Algani et al. [10] used a novel deep learning approach called Ant Colony Optimization
with Convolution Neural Network (ACO-CNN) to identify and categorize diseases. Ant
colony optimization (ACO) was used to study how well it could detect diseases in plant
leaves. Images of plants had their color, texture, and leaf arrangement removed by a CNN
classifier.

Pandian et al. [11] developed a unique Deep Convolutional Neural Network model for
identifying 42 leaf diseases across 16 plant species. They optimized the hyperparameters and
added more data to the disease detection model to boost its accuracy. Using BIM, DCGAN,
and NST, they were able to create augmented photos of leaves. In order to train the DCNN
model, 58 classes of sick and healthy plant leaves were used. To find the best settings for the
most often used hyperparameters, a random search was performed using a coarse-to-fine
approach. Finally, the suggested DCNN's performance was compared to that of industry
standard transfer learning methods.

Mahum et al. [12] created a method utilizing a refined deep learning algorithm to categorize
potato leaves into five distinct groups based on their visual characteristics. Using a pre-
existing dataset, the model that was suggested is trained (called "The Plant Village") that
contains photos of potato leaves labeled as Healthy, Normal, or Infected with Early Blight
(EB) or Late Blight (LB), respectively. To speed up the process of disease classification in
potato leaves, a pre-trained Efficient DenseNet model has been used, with the help of an
additional transition layer in DenseNet-201. Since the training data is very unbalanced, our
proposed technique benefits from the use of the reweighted cross-entropy loss function.
Overfitting is kept to a minimum in the training of tiny training sets of potato leaves because
to the thick relationships with regularization power.
Chen et al. [13] implemented a more effective plant disease recognition model using the
YOLOv5 network baseline. To begin, they employed a brand-new InvolutionBottleneck
module to cut down on parameters and calculations while simultaneously picking up spatial
information at great distances. Second, a SE module was integrated to increase the model's
responsiveness to channel characteristics. To prevent the degeneration of the loss function
"Generalized Intersection over Union" into "Intersection over Union," "Efficient Intersection
over Union" was introduced. The proposed strategies were implemented to enhance the
network model's target recognition performance.

Kawasaki et al. [14] addressed illness detection in cucumber leaves caused by viruses like
Zucchini yellow mosaic and yellow spot using a unique CNN architecture. They
demonstrated that adding more data to the mix is more helpful in recognizing performance
than simply adding more training epochs. Additionally, they dealt with numerous
augmentation procedures across a variety of topologies to boost classification accuracy.
Durmus al. [15] showed the work carried out by training AlexNet and CNNs on SqueezeNet
on the nvidia jetson platform to propose an expanded work. Alexnet's accuracy is slightly
lower than that of projects where a TITAN X GPU was employed, but this is still quite
respectable. In addition, the findings suggested that real-time experiments on crop disease
detection systems are feasible. They demonstrated that great performance in embedded
applications and transferring the model to the embedded device for use in actual use.

Jiang et al. [16] introduced the alternate method of using DL in object detectors. In addition,
the zone harboring diseases is classified and located using the features inside the bounding
box using faster region based CNNs architectures and Single Shot Multi-box detector.

Barbedo et al. [17] indicated that images of discrete spots and lesions, rather than complete
leaves, are needed for DL-based leaf disease classification. However, there are still open
questions regarding the precision with which photos can be automatically segmented into
individual lesions after the background has been removed. In addition, the mobile phone-
based detection system for plant diseases uses a compact deep CNN technology.
Additionally, CNN models can be implemented on mobile devices, increasing the usefulness
and accessibility of this technology to farmers.

Syed-Ab-Rahman et al. [18] applied a two-stage deep CNN model to identify plant diseases
and classify citrus diseases from leaf photos. A region suggestion network is used in the first
stage to suggest probable target diseased areas, and a classifier is used in the second stage to
assign the most likely aim area to the associated disease class.

Gehlot et al. [28] developed the EffiNet-TS design, which is an EfficientNetV2

Teacher/Student architecture consisting of the EffiNet-Teacher classifier, Decoder, and
EffiNet-Student classifier. Since it feeds and employs a substantial portion of a country's
population, agriculture is essential to national prosperity. One of the most difficult aspects of
increasing farmers' revenue is the presence of plant diseases. Recent advances in deep
learning have allowed for the development of highly accurate and efficient categorization
frameworks, such as efficientNet, has significantly expanded CNN's utility.

Khan et al. [29] suggested a deep learning system based on the Entropy ELM to diagnose
diseases in cucumber leaves. The proposed system is used to train one of four pre-trained
deep models with the goal of improving accuracy. Next, the Entropy-Elm method is applied
to this model in order to pick the most useful characteristics. The opposite phase, which
involves combining features from all pretrained models, is where the feature selection
method comes into play. In the end, they did classification using a combination of features
from the first two steps. The suggested framework was evaluated on a dataset of modified
cucumber leaves, and it achieved an accuracy of 97.98%.

Divyanth et al. [30] utilized SegNet, UNet, and DeepLabV3+, three semantic segmentation
models, in a two-step process. In the first phase, the complex backdrop is separated from the
leaf image so that it may be detected in the second. After evaluating the performance of each
segmentation model, they concluded that UNet was superior for stage one and that the
DeepLabV3+ model was superior for stage two. They also enhanced upon previous methods
of assessing the extent of disease by measuring the area of disease lesions.

Nandi et al. [31] implemented the VGG-16, GoogleNet, ResNet-18, MobileNet-v2, and
Efficient Net CNN models. They tried applying model quantization approaches to those three
aforementioned CNN models and discovered that GoogleNet was the most accurate and had
the smallest footprint. After quantization, the EfficientNet model reached 99% accuracy with
a manageable size.

Algani et al. [32] utilized a CNN optimized with ants (ACO-CNN). Accuracy, precision,
recall, and F1-score were all better for the ACOCNN model than for the C-GAN, CNN, and
SGD models. C-GAN has an accuracy of 99.59%, CNN of 99.89%, and SGD of 85%. With
an accuracy of 99.98%, the ACOCNN model has the highest F1 score of any currently
available model. The identification of Basal Stem Rot was a primary focus of Yong et al.'s
[33] research. They discussed deep learning and hyperspectral imaging. In this method, they
analyzed spectral changes across leaf positions by segmenting the top-down image of the
seedling into areas. To assess the role of setting of images on identification accuracy, they
segmented photos of the plant and fed them into a Mask Region-based Convolutional Neural
Network (RCNN). They used VGG16 and Mask RCNN to train their system, with VGG16
yielding the highest accuracy (94.32 %).

Ma et al. [34] utilized an attention module built into the multi-stage partial network's
backbone, they extracted multi-dimensional information from both the spatial and channel
views. To increase the breadth of information about agricultural diseases that can be extracted
from photographs of crops, they also added an area pyramid aggregating module that makes
use of dilated convolutions to the network.

Guerrero-Ibanez and Reyes-Muñoz [35] included GAN based methods for augmenting data
into the process of developing a CNN architecture for disease detection and classification in
tomato leaves. The accuracy of disease classification was improved to a level of 99.64%. To
begin with, a deep module was presented in the MMDGAN generator to enhance feature
extraction of tomato disease leaves. To regulate the overall process of image production, an
integrated attention system was then devised, which incorporated a cross attentiveness
module with a merged module. Finally, the Markov discriminator was implemented to
improve local texture similarity assessment.

Saeed et al. [36] considered the diagnosis of tomato leaf diseases by classifying photos of
healthy and unhealthy tomato leaves by using a classification system. They developed these
models by utilizing a publicly available dataset known as PlantVillage, and they achieved a
validation accuracy of 99.22%, which was the best possible score.

Ahmad et al. [37] evaluated the efficacy of five conventional deep learning models with
regard to the identification of plant diseases across a wide range of environmental variables.
The training for these models was done with photos of corn diseases taken from publicly
available sources. According to what they found, employing DenseNet169 resulted in
maximum validation accuracy (81.60%) in the field of plant disease diagnosis, representing
the peak of generalization performance. The classification of tomato leaf disease was the
topic of discussion in [38], where a strategy for fine-tuning the created CNN models was
presented. An optimization of the hyperparameters was carried out by the authors utilizing
the particle swarm optimization technique (PSO). Grid search optimization is used to get the
best results for optimizing the weights of these structures. They also suggested a triple and
quintuple ensemble model and an approach that classified the datasets called cross-validation.
They achieved the best possible classification accuracy of 99.60% by utilizing the ensembles
approach.

Francis et al. [39] outlined the automatic generation of features and the creation of prediction
systems in agriculture as an application of typical deep learning models. They placed a strong
emphasis on the significance of fine-tuning the model, transfer learning, and the segmentation
of sick areas. They began by training on a dataset consisting apple leaves, both healthy and
ill, and then they tested the performance of different MobileNet models whose depth
multipliers and resolution multipliers varied. Using both of Mobilenet and the K means
clustering method, they were able to get the greatest accuracy possible, which was 99.7%.

5. Leaf disease recognition using generative adversarial network

The field of producing synthetic images has seen the introduction of generative adversarial
networks, sometimes known as GANs, during the course of the previous half decade. In
addition, CNNs have found widespread application in the fields of disease recognition in
leaves and pest recognition. However, CNNs have been utilized in these areas, and their
efficacy as a strategy has been demonstrated. Despite this, the most significant challenge,
which is that of a restricted training dataset, has been neglected. Because of this, there is now
a problem with the data being overfit. In addition, thanks to the development of algorithms
that are based on GANs, the accuracy of predictions has improved, and the issue of excessive
data fitting due to a lack of available training data has been overcome. GANs are utilized in
[24] Goodfellow et al. primarily for the purpose of combating the issue of insufficient data.
GANs have a structure that is made up of two networks, namely a discriminator network and
a generator network. In this scenario, the generator is responsible for capturing the training
data distribution, whilst a discriminator is responsible for calculating the likelihood of
whether an image came from the program that generated it or the data set used for training. In
addition, the objective is to enhance the generator's ability to trick the discriminator, which
has been educated to recognize natural images from those that have been artificially created.
In order to employ this strategy, you will need a dedicated GANs that can generate synthetic
images. These synthetic images will then be used to train a system that can identify leaf
diseases and pests.
Arsenovic et al. [25] concluded the issue of using artificial means to generate crop photos by
employing a technique known as Generative Adversarial Networks (GANs). In addition,
throughout the course of the past few years, a number of different variations of GAN
architectures have been presented, including CGAN, DCGAN, ProGAN, and StyleGAN. In
order to more accurately reflect multimodal data production, the conditional GAN, also
known as CGAN, is utilized. According to the findings, the optimal resolution for the leaf
images that StyleGAN created was 256 pixels by 256 pixels. Because of the noisy
background, these GAN networks did not train well on field photos, which is a problem that
has not been rectified despite additional training on field images. Networks trained on GAN-
generated images performed around 1% better than networks trained just on natural images
on the test set. As an added bonus, Plant DiseaseNet (PDNet), a two-stage convolutional
neural network, was unveiled as well [25]. The initial stage (PDNet-1) of this process, which
predicted the leaf bounding box, also made use of the detector YOLOv3 and the feature
extractor Alexnet. In addition, stage two of the plant diseasenet, or PDNet-2, is comprised of
a softmax layer, a 32-layer CNN architecture, a pooling layer for global averaging, and a
fully connected 42-way layer. In terms of map scores, PDNET-1 achieves 0.9165, while
PDNET-2 achieves 93.66% accuracy in identifying agricultural diseases. It is worth noting
that GANs have huge untapped potential for creating training images automatically and are
very beneficial in solving informational difficulties shortage.

As per Zhang et al. [26], conventional deep learning-based techniques have several
drawbacks, including: (1) the need for expensive hardware and a huge amount of data to train
the models. (2) The slow inference speed of models makes them difficult to adapt to real-
world manufacturing. The third problem is that models don't generalize well enough. In light
of these challenges, this research proposes using a Tranvolution detection network equipped
with GAN modules to identify plant diseases. GAN models were first integrated into the
attention extraction module, and then GAN modules were built from scratch. After the CNN
was integrated with the updated Transformer, they proposed the Tranvolution architecture.

Chen et al. [27] developed a three-stage deep learning-based pipeline to address the issues,
which consisted of a convolutional neural netword (Faster R-CNN) for lesion detection. The
lesions on grape leaves were marked using Faster R-CNN so that a lesions dataset could be
obtained, which ResNet was utilized to identify. Leaf GAN makes better use of features to
produce grape leaf disease images with noticeable disease lesions using a decreasing-channel
producer model. The original grape disease photos are then used in conjunction with a
discriminator model that makes use of a dense connection technique and instance
normalization to produce highly accurate feature extraction results. The training process is
then stabilized by using the deep regret diminished function.

6. Leaf disease recognition based on different types of data

As the IoT spreads throughout the agricultural sector, driven by advances in digital
technology and recognizing sensors, new sensor technologies emerge and evolve in the
directions of being attached, smart, combined, and simplified.

6.1. Ground image dataset

Imaging crops from the ground up, such as with a smartphone or digital camera, is known as
"crop ground imaging." Researchers attempted to take pictures of plant leaves in the field,
which presented a number of challenges due to the presence of things like complicated
backgrounds, shadows, and varying levels of light.

For this goal, conventional machine learning methods were employed. Due to their high
accuracy in making predictions, support vector machine (SVM) models are widely utilized in
diagnostics of plant diseases as a subject of study. Hyperspectral imaging at close range for
early detection of severe drought in barley was achieved using support vector machines [40].
Data from labels and two vegetation indices were used to train the model.

Similar methods for reducing misclassification in disease detection on plant leaves, such as
physical extraction of lesions, and the combining of various support vector machine (SVM)
classifiers (color, texture, and shape attributes) [41]. Using a PCA model, we were able to
differentiate between plants in good condition and the advancement of golden potato disease
[50] based on statistical analysis of some variables. Hyperspectral pictures of diseased
potatoes at various phases of disease development were obtained. The investigation
confirmed that spectral data can be used to differentiate between plants in good condition and
those that are afflicted with illness. Similarly, the authors in [42] employed hyperspectral
images to classify the severity of grey mold infections on tomato leaves by employing K-
nearest neighbor (KNN) and the decision tree-based classifier C5.0. Using ANN with a single
hidden layer, the authors of [43] estimated the severity of three wheat illnesses. The network
achieved an 81% classification accuracy. Referring to El Massi et al. [44].
Data type ML/DL Method Crop Dataset Accuracy Reference
SVM Barley 204 images 68%
SVM Tomato 284 images 93.90%
SVM Rice 120 images 73.33%
ML PCA Potato 120 images -
KNN Tomato 212 images 92.86%
630
ANN Wheat 81%
multispectral
310
ELM Tomato hyperspectral 100
images
Ground
180
images ELM Tobacco 98%
hyperspectral
ResNet Multiple 55,038 images 99.67%
2D CNN 121,955
DL Wheat 98%
BidGRU images
Adapted
Tomato 7176 89.2
MobileNet
SSCNN Citrus 2939 99%
MobileNet Apple 334 images 73.50%
DenseNet Tomato 666 images 95.65%
EfficientNet Multiple 55,038 images 99.97%
UAV ML Hyperspectral
BPNN Tomato -
images images
Hyperspectral
CART Vine grape 94.1
images
Multispectral
ROC analysis Vine grape -
images
SLIC + SVM Soybean RGB Images 98.34
 Multispectral
Random forest Wheat 89.34
images
Multispectral
RBF Citrus 96
images
Multispectral
AdaBoost Citrus 100
images
Thermal and
SVM Olive hyperspectral 80
images
Hyperspectral
MLP Avocado 94
images
ResNet Maize RGB images 97.85
Multispectral
CNN Potato -
images
Multispectral
Net-5 Grapevine 95.86
images
CNN Maize RGB images 95.1
Hyperspectral
DL DCNN Wheat 85
images
DCGAN+
Pinus Tree RGB images -
inception
Multispectral
SegNet Grapevine -
images
Multispectral
VddNet Grapevine 93.72
images

7. Plant Disease Dataset

One such open dataset, PlantVillage, has accumulated 54309 photos of plant illnesses on its
leaves; this dataset covers 14 different types of fruit and vegetable crops, including apples,
blueberries, cherries, grapes, oranges, potatoes, peppers, pumpkins, strawberries, and
tomatoes. Corn has 12 photographs of healthy crop leaves in addition to 26 images of
diseased ones (17 fungal diseases, 4 bacterial diseases, 2 mycoses, 2 viral diseases, and 1
mite disease).
‘Plant Pathology Challenge’ for CVPR 2020-FGVC7, including 1,200 photos of apple scab,
1,399 images of cedar apple rust, 187 images of leaves with several diseases, and 865 images
of healthy apple leaves, the dataset has a total of 3,651 tagged RGB images.

8. Performance Measurement Parameters

To evaluate the efficacy of the method being presented, we have considered several metrics
such as sensitivity, specificity, dice score, accuracy, positive predicted value and area under
curve. These metrics can be computed as follows:
Performance Metrics Computation formula
Sensitivity Sn TP
TP+ FN
Specificity TN
TN + FP
Accuracy TP+TN
TN +TP+ FP+ FN
AUC ≈ 0.5 ( Sn+Sp )

Here, True positive (TP), false positive (FP), false negative (FN), and true negative (TN)
stand for the opposite of each other. Sensitivity ( Sn ) is the true positive rate, Specificity ( S p )
specificity is proportion of true negative and false positive, accuracy represents the rate of
actual prediction, false positive represents the incorrect positive predictions, false negative
denotes incorrect negative prediction. Finally, AUC is computed which is approximately near
to half of the sum of sensitivity and specificity.

9. Conclusion

In this paper, we presented a foundational understanding of deep learning and give a

thorough evaluation of current research into the application of deep learning to the problem
of plant leaf disease recognition. Assuming a sufficient amount of training data is made
accessible, deep learning methods may accurately detect leaf-borne diseases. Hyper-spectral
imaging and small-sample disease detection in plant leaves have been discussed, as have the
benefits of collecting huge databases with high variation, improving data quality, learning
from experience, and visualizing convolutional neural network (CNN) activation maps to
improve classification accuracy. There are, however, a few shortcomings. Most DL
frameworks proposed in the literature perform well on their own datasets, but not so well on
other datasets. This indicates that the model is not very robust. Improved DL models are
needed to accommodate the wide range of disease datasets.

The PlantVillage dataset was utilized in many studies to measure the efficacy of DL models.
Despite the large number of photos depicting various plant diseases, this collection is still
useful. A comprehensive database of plant diseases under natural settings is thus anticipated.
Even seasoned professionals have trouble pinpointing the precise location of invisible disease
symptoms and defining pure invisible illness pixels, making it challenging to generate the
annotated information necessary for early plant disease detection.
References

1. Ramesh, S., Hebbar, R., Niveditha, M., Pooja, R., Shashank, N., & Vinod, P. V.
(2018, April). Plant disease detection using machine learning. In 2018 International
conference on design innovations for 3Cs compute communicate control
(ICDI3C) (pp. 41-45). IEEE.
2. Harakannanavar, S. S., Rudagi, J. M., Puranikmath, V. I., Siddiqua, A., &
Pramodhini, R. (2022). Plant leaf disease detection using computer vision and
machine learning algorithms. Global Transitions Proceedings, 3(1), 305-310.
3. Chug, A., Bhatia, A., Singh, A. P., & Singh, D. (2022). A novel framework for image-
based plant disease detection using hybrid deep learning approach. Soft Computing, 1-
26.
4. Saleem, M. H., Potgieter, J., & Arif, K. M. (2022). A performance-optimized deep
learning-based plant disease detection approach for horticultural crops of new
zealand. IEEE Access, 10, 89798-89822.
5. Nandhini, M., Kala, K. U., Thangadarshini, M., & Verma, S. M. (2022). Deep
Learning model of sequential image classifier for crop disease detection in plantain
tree cultivation. Computers and Electronics in Agriculture, 197, 106915.
6. Arun, R. A., & Umamaheswari, S. (2023). Effective multi-crop disease detection
using pruned complete concatenated deep learning model. Expert Systems with
Applications, 213, 118905.
7. Chen, J., Zeb, A., Nanehkaran, Y. A., & Zhang, D. (2022). Stacking ensemble model
of deep learning for plant disease recognition. Journal of Ambient Intelligence and
Humanized Computing, 1-14.
8. Hassan, S. M., & Maji, A. K. (2022). Plant disease identification using a novel
convolutional neural network. IEEE Access, 10, 5390-5401.
9. Thakur, P. S., Sheorey, T., & Ojha, A. (2023). VGG-ICNN: A Lightweight CNN
model for crop disease identification. Multimedia Tools and Applications, 82(1), 497-
520.
10. Abd Algani, Y. M., Caro, O. J. M., Bravo, L. M. R., Kaur, C., Al Ansari, M. S., &
Bala, B. K. (2023). Leaf disease identification and classification using optimized deep
learning. Measurement: Sensors, 25, 100643.
11. Pandian, J. A., Kumar, V. D., Geman, O., Hnatiuc, M., Arif, M., & Kanchanadevi, K.
(2022). Plant disease detection using deep convolutional neural network. Applied
Sciences, 12(14), 6982.
12. Mahum, R., Munir, H., Mughal, Z. U. N., Awais, M., Sher Khan, F., Saqlain, M., ... &
Tlili, I. (2023). A novel framework for potato leaf disease detection using an efficient
deep learning model. Human and Ecological Risk Assessment: An International
Journal, 29(2), 303-326.
13. Chen, Z., Wu, R., Lin, Y., Li, C., Chen, S., Yuan, Z., ... & Zou, X. (2022). Plant
disease recognition model based on improved YOLOv5. Agronomy, 12(2), 365.
14. [14] Y. Kawasaki, H. Uga, S. Kagiwada, H. Iyatomi, Basic study of automated
diagnosis of viral plant diseases using convolutional neural networks, in: International
symposium on visual computing, Springer, 2015, pp. 638–645.
15. H. Durmus¸, E. O. Gunes¸, M. Kırcı, Disease detection on the leaves ¨ of the tomato
plants by using deep learning, in: 2017 6th International Conference on Agro-
Geoinformatics, IEEE, 2017, pp. 1–5.
16. P. Jiang, Y. Chen, B. Liu, D. He, C. Liang, Real-time detection of apple leaf diseases
using deep learning approach based on improved convolutional neural networks,
IEEE Access 7 (2019) 59069–59080.
17. J. G. A. Barbedo, Plant disease identification from individual lesions and spots using
deep learning, Biosystems Engineering 180 (2019) 96–107.
18. Syed-Ab-Rahman, S. F., Hesamian, M. H., & Prasad, M. (2022). Citrus disease
detection and classification using end-to-end anchor-based deep learning model.
Applied Intelligence, 52(1), 927-938.
19. Dake, W.; Chengwei, M. The support vector machine (SVM) based near-infrared
spectrum recognition of leaves infected by the leafminers. In Proceedings of the First
International Conference on Innovative Computing, Information and Control-Volume
I (ICICIC’06), Beijing, China, 30 August–1 September 2006; Volume 3, pp. 448–451
20. Duarte-Carvajalino, J.M.; Alzate, D.F.; Ramirez, A.A.; Santa-Sepulveda, J.D.;
Fajardo-Rojas, A.E.; Soto-Suárez, M. Evaluating late blight severity in potato crops
using unmanned aerial vehicles and machine learning algorithms. Remote Sens. 2018,
10, 1513.
21. Raza, S.E.A.; Prince, G.; Clarkson, J.P.; Rajpoot, N.M. Automatic detection of
diseased tomato plants using thermal and stereo visible light images. PLoS ONE
2015, 10, e0123262
22. Mokhtar, U.; Ali, M.A.; Hassanien, A.E.; Hefny, H. Identifying two of tomatoes leaf
viruses using support vector machine. In Information Systems Design and Intelligent
Applications; Springer: Berlin/Heidelberg, Germany, 2015; pp. 771–782
23. Brahimi, M.; Boukhalfa, K.; Moussaoui, A. Deep learning for tomato diseases:
Classification and symptoms visualization. Appl.Artif. Intell. 2017, 31, 299–315.
24. I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A.
Courville, Y. Bengio, Generative adversarial nets, Advances in neural information
processing systems 27 (2014).
25. M. Arsenovic, M. Karanovic, S. Sladojevic, A. Anderla, D. Stefanovic, Solving
current limitations of deep learning based approaches for plant disease detection,
Symmetry 11 (7) (2019) 939
26. Zhang, Y., Wa, S., Zhang, L., & Lv, C. (2022). Automatic plant disease detection
based on tranvolution detection network with GAN modules using leaf images.
Frontiers in Plant Science, 13, 875693.
27. Chen, Y., & Wu, Q. (2023). Grape leaf disease identification with sparse data via
generative adversarial networks and convolutional neural networks. Precision
Agriculture, 24(1), 235-253.
28. Gehlot, M., & Gandhi, G. C. (2023). “EffiNet-TS”: A deep interpretable architecture
using EfficientNet for plant disease detection and visualization. Journal of Plant
Diseases and Protection, 130(2), 413-430.
29. M. A. Khan et al., “Cucumber Leaf Diseases Recognition Using Multi Level Deep
Entropy-ELM Feature Selection,” Applied Sciences, vol. 12, no. 2, p. 593, Jan. 2022,
doi: 10.3390/app12020593.
30. L. G. Divyanth, A. Ahmad, and D. Saraswat, “A twostage deep-learning based
segmentation model for crop disease quantification based on corn field imagery,”
Smart Agricultural Technology, vol. 3, p. 100108, Aug. 2022, doi:
10.1016/j.atech.2022.100108
31. R. N. Nandi, A. H. Palash, N. Siddique, and M. G. Zilani, “Device-friendly Guava
fruit and leaf disease detection using deep learning,” arXiv (Cornell University), Sep.
2022, doi: 10.48550/arxiv.2209.12557.
32. Y. M. A. Algani, O. J. M. Caro, L. M. Robladillo-Bravo, C. Kaur, M. S. AlAnsari,
and B. K. Bala, “Leaf disease identification and classification using optimized deep
learning,” Measurement: Sensors, vol. 25, p. 100643, Dec. 2022, doi:
10.1016/j.measen.2022.100643
33. L. Z. Yong, S. K. Bejo, M. Jahari, and F. M. Muharam, “Automatic Disease
Detection of Basal Stem Rot Using Deep Learning and Hyperspectral Imaging,”
Agriculture, vol. 13, no. 1, p. 69, Dec. 2022, doi: 10.3390/agriculture13010069
34. W. Ma et al., “Crop Disease Detection against Complex Background Based on
Improved Atrous Spatial Pyramid Pooling,” Electronics, vol. 12, no. 1, p. 216, Jan.
2023, doi: 10.3390/electronics12010216
35. J. A. Guerrero-Ibanez and A. Reyes-Muñoz, “Monitoring Tomato Leaf Disease
through Convolutional Neural Networks,” Electronics, vol. 12, no. 1, p. 229, Jan.
2023, doi: 10.3390/electronics12010229
36. A. Saeed, A. A. Abdel-Aziz, A. Mossad, M. A. Abdelhamid, A. Y. Alkhaled, and M.
Mayhoub, “Smart Detection of Tomato Leaf Diseases Using Transfer Learning-Based
Convolutional Neural Networks,” Agriculture, Jan. 2023, doi:
10.3390/agriculture13010139
37. A. Ahmad, A. E. Gamal, and D. Saraswat, “Toward Generalization of Deep Learning-
Based Plant Disease Identification Under Controlled and Field Conditions,” IEEE
Access, vol. 11, pp. 9042–9057, Jan. 2023, doi: 10.1109/access.2023.3240100
38. H. Ulutaş and V. Aslantaş, “Design of Efficient Methods for the Detection of Tomato
Leaf Disease Utilizing Proposed Ensemble CNN Model,” Electronics, vol. 12, no. 4,
p. 827, Feb. 2023, doi: 10.3390/electronics12040827
39. M. Francis, K. S. M. Anbananthen, D. Chelliah, S. Kannan, S. Subbiah, and J.
Krishnan, “Smart Farm-Care using a Deep Learning Model on Mobile Phones,”
Emerging Science Journal, vol. 7, no. 2, pp. 480–497, Feb. 2023, doi: 10.28991/esj-
2023-07-02-013.
40. Behmann, J.; Steinrücken, J.; Plümer, L. Detection of early plant stress responses in
hyperspectral images. ISPRS J. Photogramm. Remote Sens. 2014, 93, 98–111
41. Es-Saady, Y.; El Massi, I.; El Yassa, M.; Mammass, D.; Benazoun, A. Automatic
recognition of plant leaves diseases based on serial combination of two SVM
classifiers. In Proceedings of the 2016 International Conference on Electrical and
Information Technologies (ICEIT), Tangiers, Morocco, 4–7 May 2016; pp. 561–566.
42. Xie, C.; Yang, C.; He, Y. Hyperspectral imaging for classification of healthy and gray
mold diseased tomato leaves with different infection severities. Comput. Electron.
Agric. 2017, 135, 154–162.
43. Bebronne, R.; Carlier, A.; Meurs, R.; Leemans, V.; Vermeulen, P.; Dumont, B.;
Mercatoris, B. In-field proximal sensing of septoria tritici blotch, stripe rust and
brown rust in winter wheat by means of reflectance and textural features from
multispectral imagery. Biosyst. Eng. 2020, 197, 257–269
44. . El Massi, I.; Es-saady, Y.; El Yassa, M.; Mammass, D.; Benazoun, A. Hybrid
combination of multiple svm classifiers for automatic recognition of the damages and
symptoms on plant leaves. In Image and Signal Processing, ICISP 2016; Lecture
Notes in Computer Science; Springer: Cham, Switzerland; Volume 9680