Multimedia Tools and Applications (2023) 82:36091–36117

[Link]

Automatic early detection of rice leaf diseases

using hybrid deep learning and machine learning
methods

Vikram Rajpoot 1 & Akhilesh Tiwari 1 & Anand Singh Jalal 2

Received: 4 April 2022 / Revised: 1 September 2022 / Accepted: 22 February 2023 /
Published online: 15 March 2023
# The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2023

Abstract
Plant leaf disease detection is critical for long-term agricultural viability. Numerous
Artificial Intelligence (AI) and Machine Learning (ML) technologies have been imple-
mented for detecting rice diseases. However, such methods failed to identify or have slow
recognition causing severe output loss. Therefore, an advanced and precise detection
method has become necessary to overcome this issue. This study analyzes plant diseases
that affect rice, comprising three different forms of diseases. Bacterial leaf blight, Brown
spot, and Leaf smut are three of the six diseases that can affect rice plants. In the proposed
approach a VGG-16 transfer learning with Faster R-CNN deep architecture is used to
extract features. After completing the transfer learning step, the gathered characteristics
are categorized using the random forest method. The random forest classifier divided the
radish field into three distinct regions. The images of rice plant leaves are taken from UCI
Machine Learning Repository. The proposed approach obtains an average predicting
accuracy of 97.3% for rice disease imagery class prediction. The extensive experiment
outcomes demonstrate the suggested technique’s validity, so it effectively detects rice
diseases.

Keywords Plant Disease . Rice leaf disease . Machine learning . Deep Learning . Faster R-CNN .
VGG-16 . Random Forest

* Anand Singh Jalal

asjalal@[Link]

Vikram Rajpoot
vikramraj@[Link]
Akhilesh Tiwari
atiwari@[Link]

1
Department of Information Technology, Madhav Institute of Technology & Science, Gwalior,
Madhya Pradesh, India
2
Department of Computer Engineering and Applications, GLA University, Mathura, Uttar Pradesh,
India
36092 Multimedia Tools and Applications (2023) 82:36091–36117

1 Introduction

The agriculture industry has begun exploring new food production methods due to the growing
population. The primary source of economic development has been agriculture in India. The
farmers select suitable crops to build based on the soil type, location, weather, and economic
value. Rice plant diseases are critical issues for agriculture, economic and societal losses in the
agricultural field’s future growth. This enables researchers to strive for new high-productivity
technologies which are efficient and accurate. To collect information and data for the best
decision on farm output, farmers should use precision agriculture. Precision agriculture is a
state-of-the-art technology for increasing farm output. Economic growth in agriculture can be
achieved with these advanced technologies. Precision agriculture can be used in various uses,
including plant pest detection, weed detection, crop yield or plant disease recognition, etc.
Farmers use pesticides to prevent pests and diseases and increase crop yields. Crop disease
creates problems for farmers and agricultural industries with poor production and economic
losses. Identification and severity of disease are also required. Researchers have been inter-
ested in detecting plant diseases for the last ten years using image processing methods. For
effective farming systems, disease identification in crops is the most important. In a wide range
of crops, several disease identification, detection, and quantifying models have been proposed
and implemented. Because rice feeds approximately 50% of the worldwide population,
detecting rice leaf disease is critical for global agriculture. A farmer usually identifies signs
of disease in plants with naked-eye observations, which is essential to monitor continuously.
However, this method is costly in large plantations and may often be less effective. Even, in
many cases, farmers may need to show the specimen to experts, which takes time and makes it
more expensive [48].
In every plant, bacteria and fungi are causing diseases. For rice plants, Brown spots, Leaf
blasts, Bacterial leaf blight, Leaf smut, and Sheath blight are diseases [23]. The characteristics
of each disease are unique. The symptoms of the disease differ in form, size, and color. Some
cases may be the same color but different forms, while others have different colors but various
forms. Farmers are often confused and can not make the right decision on pesticide use. It is
one way to capture images of the infected leaves and get information about the disease to
eliminate the loss of the seed from the infection of the disease. Cameras may be used
periodically to capture images at some distances on the farm as an automated solution to this
problem [37].
Researchers have worked on such topics, and their results varied depending on the
approaches used. Since conventional farming requires manual data collection, volatile climate
conditions, a smattering of pesticides on crops, disease diagnosis, etc., Machine Learning
(ML) methods are increasing enormously in real-time demands. Concerning the current
situation in conventional agriculture, there has been a desperate need for data predicted in
crop production that allows farmers to know about their issues in real-time & also to respond
accordingly [35]. Real-time applications using supervised and unsupervised approaches in
machine learning have become increasingly popular. In a crop disease, low productivity leads,
and, on the other hand, the farmers suffer huge losses. Thus, early detection of diseases will
support farmers in creating necessary actions [1].
Time complexity is one of the significant issues when using unsupervised disease-detecting
models. The high accuracy value of the k-mean [52] must be computed for the indexed period.
A further significant issue is the vulnerability of segmentation toward this evaluation of region
of interest (ROI). Researchers have tried to hybridize unsupervised models with those
Multimedia Tools and Applications (2023) 82:36091–36117 36093

mathematical measures [3, 45] and AI algorithms to improve their results. In the field of AI
application, [9] introduced a novel method for approaches called fb-kNN, which has been
based on the study of similarities of illnesses with sequences in the human body. This
algorithm was later applied in Healthcare 4.0 to suggest diagnosis and therapy. The recom-
mendation system assists these operations by utilizing deep data mining and obtaining the data
of other hospitals to identify areas with low vaccination rates [8]. According to the study of
[10], Air pollution has a negative impact on environmental quality and global climate change,
which requires almost every country to make policies to improve air quality that are impacted
to agriculture. The problem of air pollution has become an important issue that urgently needs
to be solved and has attracted the attention of both the government and society. In real-time
disease detection, a trend was also found to use DL models to address quite complicated
problems. One study [5] obtained an average accuracy of 98.42% by using the GoogLeNet
architecture in CNN to identify apple-leaf diseases using a single historical plant village
dataset. A new model based on MobileNets, based on two coevolutionary stages, is suggested
by another survey [29]. The first is to introduce a depth-separable convolution in a single
batch. Secondly, point-by-point convergence includes facts deeply. This novel framework
aimed to use the mobile app, which reduces the computational delay problem by 6 times less
than the MobileNet standard.
The suggested discussions focused only on the existing plant disease literature. There have
been no scalable and available adoption models to classify and forecast highly complex plant
disease problems. Due to inadequate knowledge or limited availability of trained experts, it is
challenging for all the growers to detect diseases and fail to identify or slow recognition. No
image enhancement methods have been employed to derive detailed classification features.
The overall performance results of the analysis are not forecasted because of the more
excellent training set and testing sets of data. This paper discusses the number of research
gaps with their limitations in the existing systems and methods. An advanced and precise
detection method can solve such issues and limitations. It has motivated that it was possible to
increase productivity from the declining agricultural land, which would be capable of feeding
India’s billion-plus populace in the future. As per early findings, 42% of crop yields are lost
due to the rising rate of loss attributed to plant leaf diseases. This plant-leaf disease detection
methodology can be required to evaluate disease from the image data to resolve such a
significant issue. So this work’s primary motivation is effectively detecting rice leaf disease.
Implementing AI (Artificial Intelligence) and efficient machine vision technologies have
provided a specific and productive method for detecting rice diseases. This can be done by
incorporating the advanced classification model to merge with DL’s AI system.
The proposed system can predict the early disease detection of rice plants. Different kinds
of diseases in rice leaves are identified as per geographical distribution and favorable weather
conditions. Table 1 represents all these diseases with their symptoms. Decision tree leveraging
factors like temperature, moisture levels, etc., based on previous year’s information and via
sensors. Diseases in crop yield are reduced productivity, contributing to a tremendous loss for
producers. Consequently, early disease diagnosis would be advantageous for farmers so that
appropriate steps can be taken to decrease the growth of unhealthy grains.
The key contributions are considered as following for this work:

& This work aims to implement machine learning & image processing concepts to solve
automatic rice plant disease detection and classification, which is India’s one most
important food.
36094 Multimedia Tools and Applications (2023) 82:36091–36117

Table 1 Various disease identification for rice leaves

S.N Leaf Image Symptoms Disease Geographical Favorable
o Diagnosis Distribution weather
condition

1 a grayish-brown Rice Blast approximately Temperate,

lesion girdles the [14] 85 countries subtropical, and
infected part throughout the tropical.
world

2 Leaves that have Bacterial Indonesia, the tropical and

been infected will leaf blight Philippines, In temperate
curl and become [15] dia, and China environments
a grayish-green
color. The illness
causes the leaves
to wilt and
transform from
green to a straw-
like yellow as it
spreads.
3 rice crop that has Sheath Asia temperate,
been tillered up blight [20] subtropical, and
to the milk phase tropical

4 The fungus Leaf smut Tamil Nadu in high relative

produces slightly [36] India and humidity (89 to
raised, angular, many 95%) and no
black spots (sori) countries of rainfall
on both sides of the world
the leaves

5 The Sheath rice-growing Leaf-sheaths at

lesions appear as Rot [45] areas room temp and
oblong or of the world, for ten months in
somewhat including the leaf sheaths
irregular spots Bangladesh, of the plant when
with brown Cameroon, it was growing in
margins and grey India, Korea, the field.
centers which Japan, Peru,
enlarge- and Philippines,
cover most of the Southeast
leaf sheath Asia, Taiwan,
Thailand,
Venezuela,
Vietnam, and
the USA.
6 Produced Brown In High humidity
pinhead size spot [49] India,Coimbat (>92.5%), leaf w
spots, which ore, Tanjavur, etness, and
gradually Gudalur, temperature (24
enlarged to form Bavanisagar to 30°C)
an oval-shaped and
spot. Dharapuram.
Multimedia Tools and Applications (2023) 82:36091–36117 36095

& The fundamental purpose is to detect diseases via the ML(Machine Learning) approach in
the rice plant at its early growth stage. This can reduce the number of unwanted and
unhealthy seeds from developing into healthy ones.
& To overcome the search for unhealthy grain among the whole grain issues, we have
proposed a hybrid model including faster R-CNN with VGG-16 and a Random Forest
classifier.
& Early diagnosis of a disease like Brown spots, Leaf blasts, Bacterial leaf blight, Leaf smut,
and Sheath blight for rice plants would be advantageous for farmers so that appropriate
steps are taken to decrease the growth of unhealthy grains.
& Compared with existing approaches helps to validate the results.

However, several previous studies have already used state-of-the-art methods for detecting as
well as classifying rice plant leaves diseases. Regrettably, these studies mostly use specific
machine learning or deep learning techniques or integration of various machine learning or
deep learning techniques. Furthermore, there has been very little research involved in the
combinations of machine learning with deep transfer learning so it can enjoy the benefits of
both methods in this model. The novelty of this work is to propose a hybrid model of deep
learning and machine learning. In contrast, the combination of two transfer learning tech-
niques, Faster R-CNN and VGG-16, are used for feature extraction from rice leaf images.
These features have input to the Random forest machine learning classifier to train the model
and perform the classification of rice leaf diseases. In this case, deep transfer learning helps
in providing improved feature extraction with a hyperparameter setup, and the use of a
machine learning model assists in giving adequate classifications and predictions of leaf
diseases. This model is helpful for the early prediction of diseases in rice leaves, including
Brown spots, Leaf blasts, Bacterial leaf blight, Leaf smut, and Sheath blight; based on
prediction results, can use some remedies to overcome these diseases of rice leaf and the
growth of rice plants can be increased. This helps to validate the findings by checking both
the uniqueness of the approach and the results of using it against other ways that are already
considered to be state of the art.

1.1 Motivation

The agricultural sector has been an essential driver of economic growth. The agrarian sector
has been looking for new ways to increase food production due to the growing population,
climate change, and political uncertainty. Researchers look for new, high-productivity, prac-
tical, and detailed technologies. Plus, half of the diseases are spreading worldwide due to
unhealthy grains, which increases the interest of researchers in finding the various methods for
detecting several parameters of agriculture.
Farming relies heavily on environmental factors. In addition to this, tremendous energy
is used for farming. Numerous farmers’ children tend to choose other career opportunities
due to the arduous work in agriculture. It has motivated that it is possible to increase
productivity from the declining agricultural land, which would be capable of feeding
India’s billion-plus populace in the future. As per early findings, 42% of crop yields are
lost due to the rising rate of loss attributed to plant leaf diseases. This plant-leaf disease
detection methodology can be required to evaluate disease from the image data to resolve
such a significant issue. So this research’s fundamental motivation is to detect rice leaf
diseases effectively.
36096 Multimedia Tools and Applications (2023) 82:36091–36117

1.2 Objectives

& The fundamental purpose of this work is to detect diseases via the ML(Machine Learning)
approach in the rice plant at the early stage of its growth so that this can reduce the
unwanted and unhealthy seeds from developing with the healthy ones and after mixing it is
challenging to search for unhealthy grain among the whole grain.
& Establish easy access to information, expertise, and genetic information to enhance variety
with technological development.
& Organization of applied research establishes varieties, including innovations specific to
locations
& Diffusion of technology and capacity evolution.

This paper is organized further in the following structure. Section 2 defines the related work in
defining and classifying leaf diseases. Next, Section 3 discussed some existing issues in the
previous study and detailed the proposed methodology. The experimental result and their
evaluation have also been discussed in Section 4. At last, conclude this paper in Section 5 with
future discussions.

2 Literature survey

Plant diseases are adverse circumstances that cause a considerable drop in the grade and the
proportion of crops that may be produced. This approach is sometimes inaccurate and may
take a significant amount of time. Yet, it is frequently used by skilled producers and biologists
to inspect plants for signs of disease visually. The authors of this work, Singh et al. [50], aimed
to accomplish the objective of inventing and constructing an intelligent classification system
for leaf diseases. To do so, they used available artificial intelligence and computer vision
methods. During the first phase, data augmentation is carried out on the images included in the
PlantVillage data set (for apple, maize, potato, tomato, and rice plants). A CNN is used to
retrieve the in-depth features from those images. A Bayesian optimal SVM classifier is used to
classify such characteristics, and the outcomes are measured in precision, sensitivity, f-score,
and accuracy. The approaches described above will make it possible for producers across the
globe to take proper action to protect their harvests from suffering irreparable harm, which will
save both the world as a whole and the producers themselves from the possibility of a financial
catastrophe. Preparing the data set images is where the second step of the approach begins.
After that, the textured and color aspects of the images are extracted using a histogram of
oriented gradient (HoG), a general linear combination model (GLCM), and color moments. In
this case, the three different kinds of features—color, texture, and in-depth features—are
merged to generate a new kind of feature known as hybrid features. First, the binary particle
swarm optimization algorithm was used to select all these hybrid features to obtain the
simulations’ outcomes. A random forest classifier was used for the classification of the data.
Binary particle swarm optimization is an essential component of the hybrid feature selection
process; this algorithm aims to provide an appropriate output while using the fewest possible
features. The comparative study of the two methods is shown below, using the assessment
parameters discussed in this paragraph. Then Pushkara Sharma et al. in [46]developed AI-
based Automated plant leaves disease classification & detection for fast and easy monitoring.
They then classified the disease and administered solutions to recover it.
Multimedia Tools and Applications (2023) 82:36091–36117 36097

At the moment, the various deep neural network learning algorithms have accomplished a
great deal for classifying hyperspectral images (HSIs), with the CNN being used by most of
these methods. HSI data are characterized by multidimensional nature, correlations, non-
linearity, and a significant volume of data. In addition, there is a substantial volume of data.
Thus, it is especially vital to recover deeper characteristics in HSIs by lowering the dimen-
sionalities, as this helps enhance classification in both the spectrum and the spatial domains. In
their paper [11], Bhatti et al. presented a spatial-spectral HSI classification algorithm called
local similarity projection Gabor filtering (LSPGF). This algorithm combines a reduced-
dimensional CNN based on LSP with an algorithm for 2-dimensional Gabor filtering. The
following research [12], conducted in Karachi, discovered themes of variety in the city’s
wayside plants. It provides the basis for future planning for plant protection, such as the
protection of plant species, the maintenance of plant habitats, and the coordination of plant
management in Karachi. In [4], enriched leaf disease detection datasets were produced through
fundamental image modification as well as image expansion strategies based on DL (deep
learning) like image rotation, resizing, flipping, hue transition, PCA dye improvement, noise
insertion, GANs (generative adversarial networks) as well as NST (neural style transfer
methodology). The efficiency of the methods used for data extension was analyzed by
leveraging state-of-the-art transform learning methods like VGG16, ResNet, or InceptionV3.
[19]proposed a technique that uses Tensorflow, an object-tracking API, to diagnose diseases
from leaf disease images, and the framework was equipped with a faster RCNN process [7].
Also, the precision rate is measured. Lu[33]suggested a modern way of detecting rice diseases
focused on deep CNNs techniques. And use a 500 natural pictures dataset of rice leaves &
stems from the experimental field of rice. The proposed CNN-based model has obtained an
overall 95.48% accuracy under the 10-fold cross-validation approach.
On the other hand, Deep CNNs usually demand a substantial quantity of training data, are
incapable of being translated, and call for many parameters to be set and adjusted. In this
research, a very efficient framework can be used during the feature extraction process to identify
numerous leaf diseases found in plants and fruits. It uses a deep transfer learning model tailored
specifically for this use. In conclusion, Saberi, Anari, and Maryam [42] extracted characteristics
using model engineering (ME). To improve feature discrimination and computing performance,
several SVM models are used. Then continue this work, Hasan [27] incorporated the SVM
classification model to merge with an AI system called deep CNN. TL method has been applied
to develop a model that was being proposed. Afterward, 1080 images of nine distinct rice leaf
diseases from existing data collection are used to retrain their proposed framework.
Then, the SVM classification algorithm is equipped with the characteristics derived from
the DCNN method. The suggested framework expertly described and categorized rice diseases
of nine varieties and achieved an accuracy of 97.5%. Govardhan and Veena [26] introduced an
automated method for detecting and classifying plant pathogens using ML and IP (image
processing) techniques. The extracting features procedure is implemented for the images to
develop the model. The experiment file includes the unseen images used to verify device
quality in plant disease diagnosis. The system’s overall reliability is 95%. Later, classification
strategies were hybridized based on K-mean s cluster analysis, CNN, and SVM for plant
infection/target contamination in a Rice crop. The dataset training initially included ten image
collections, which were enhanced to more than 50 to get more reliability than 95%. In[13]-
collected, the infected leaves images were captured and preprocessed these collected images.
Once these image datasets have been preprocessed, features are extracted and ranked based on
specific disease-impacted leaves. Next, SVM has been applied to classify plant diseases and
36098 Multimedia Tools and Applications (2023) 82:36091–36117

suggest the proper pesticide with the correct dosage. Also, it has recommended some preven-
tion methods to prevent the plants from further issues. This process identifies the rice leaves
diseases like Brown spots, Bacterial leaf blight, & leaf Smut.
Conventional procedures for leaf disease identification are frequently wrong, time-consum-
ing, and costly; machine-driven disease diagnostic systems may be able to minimize some of
these drawbacks. Rice leaf disease detection technologies that are helped by computer
technology are becoming more common in today’s society. However, the effectiveness of
the system and the ways in which it can be used are hindered by several limitations. These
include prominent image backgrounds, vague symptoms’ edges, differences in the weather at
the time of capturing images, a lack of actual field rice leaf image data, variations in symptoms
that are caused by the same infection, multiple infections that produce symptoms similar, and
an inefficient real-time framework. Ramesh and Vydeki [39] presented an ML methodology
for recognizing disease symptoms in rice plants. Images of infected leaves containing safe and
blast diseases are captured for the proposed method. The characteristics are derived from
segments of the rice leaf which are affected by hygiene and infection. The overall collection of
data comprises 300 pictures. Such images are analyzed using the proposed methodology, and
the leaf is either labeled as infectious or safe. The simulation outcomes include 99% accuracy
for the blast-infected pictures with 100% precision for the usual images during the training
process. In [2]described, an ML algorithm for rice leaves disease detection. Some of the most
prevalent chronic diseases of rice plants are found, such as leaf smut, bacterial leaf blight, &
brown spot diseases. The input was clear pictures of the rice plant leaves impacted in white
terms. The dataset was trained with numerous techniques, including KNN, J48, Naive Bayes
& Logistic Regression, following the required preprocessing. After 10-fold cross-validation for
the decision tree classifier, the precision of the testing dataset set is around 97% accuracy.
A comparison of different approaches for different leaf diseases done by several authors in
different durations is shown in Table 2. This table indicated a lot of work using different types
of techniques and metrics that can identify the disease or infection.

3 Proposed work

This section discusses the number of issues in the existing system. Then, to overcome these
issues, proposed a hybrid model including faster R-CNN with VGG-16 and a Random Forest
classifier. A detailed view of this proposed methodology has been given in this section. İt also
describes the proposed algorithm and flow diagram of the proposed hybrid model.

3.1 Problems in the existing system

The suggested discussions focused only on the existing plant disease literature.

& Systems do not have an improved multi-convolution analysis of the data.

& There have been no scalable and available adoption models to classify and forecast highly
complex plant disease problems.
& No image enhancement methods have been employed to derive detailed classification
features. The overall performance results of the analysis are not forecasted because of the
more excellent training set and testing sets of data.
& Failure to identify or slow recognition.
Table 2 Comparative table of different approaches for different leaves disease

Author’s & year of publication Research paper name Type of technique/Performance metrics/ Disease / infection
work

Y. Lu, S. Yi, N. Zeng, Y. Liu, Y. Zhang, Identification of rice diseases using deep rice diseases identification technique CNNs are trained to the identification
(2017) convolutional neural networks [33] dependent on the deep CNNs often typical rice diseases
A. Johannes, A. Picon, A. Alvarez-Gila, Automatic plant disease diagnosis using mobile generalized multi-disease identification wheat diseases septoria, rust & tan
J. Echazarra, S. capture devices applied to a wheat use case [28] method with image-based automated spot
Rodriguez-Vaamonde, identification systems.
A.D. Navajas, A.O. Barredo, (2017)
[Link], D. Dinh,[Link] (2017) An IoT Environmental Data Collection System IoT system with cloud storage & a machine prevention of fungal disease spread
for Fungal Detection in Crop Fields [53] learning technique using SVMregression
K.P. Ferentinos, (2018) Deep learning models for plant disease detection convolutional neural network models disease detection and diagnosis using
and diagnosis [22] simple leaves images
S. Ramesh, D. Vydeki (2018) Rice Blast Disease Detection and Classification machine-learning algorithm find the symptoms of the disease in
using Machine Learning Algorithm [39] the rice plant
J. Chopda, S. Nakum, V. Nakrani (2018) Cotton Crop Disease Detection using Decision Decision Tree predict cotton crop diseases
Tree Classifier [18]
Multimedia Tools and Applications (2023) 82:36091–36117

R. Shima, [Link] (2018) Plant Disease Detection Using Machine Learning Random forest recognize abnormalities in papaya
[34] leaves
BS. Kusumo, A. Heryana, [Link], H.F. Machine Learning-based for Automatic Four machine learning algorithms Automatic Detection of Corn-Plant
Pardede (2018) Detection of Corn-Plant Diseases Using Image Diseases
Processing [31]
W.S. Kim, D.H. Lee, Y.J. Kim (2019) Machine vision-based- automatic disease symptoms Convolution neural network model with automatically crop monitoring for
detection of onion downy mildew [30] An Artificial neural network (ANN) detecting the disease symptom
A. Barbosaa, R. Trevisanb, N. Modeling yield response to crop management Convolutional Neural Network (CNN) the model produces a response to
Hovakimyana, N.F. Martinb (2019) using convolutional neural networks [6] manage the nutrient & seed rate
T. chokey, S. Jain (2019) Quality Assessment of Crops using Machine supervised machine learning techniques detect the disease in the Maize leaf
Learning Techniques [17]
U. Shruthi, V. Nagaveni, B.k. A Review on Machine Learning Classification Five types of machine learning recognition of plant disease
Raghavendra (2019) Technique for Plant Disease Detection [47] classification techniques
K. Ahmed, T. Rahman Shahidi, Rice Leaf Disease Detection Using Machine Machine learning approaches rice plant diseases like bacterial leaf
S. Md. I. Alam, S. Mome (2019) Learning Techniques [2] blight, leaf smut, &brown spots
diseases detected
36099
Table 2 (continued)
36100

Author’s & year of publication Research paper name Type of technique/Performance metrics/ Disease / infection
work

G. Verma, C. Taluja, A.K. Vision-Based Detection and Classification of Disease K-means, SVM & CNN classification techniques for rice crop
Saxena (2019) on Rice Crops Using Convolutional Neural disease/infection
Network [54]
S.H. Leea, H. Goëaua, P. A new perspective on plants diseases classification different transfer learning mechanismscapture the visual symptoms of plant
Bonneta, A. Jolyc (2020) based on deep learning [32] diseases
P. Sharma, P. Hans, S.C. Gupta Classification of Plant Leaf Diseases Using Machine Using ML& Image Preprocessing methods artificial intelligence-based automatic
(2020) Learning And Image Preprocessing Techniques plant leaf disease detection
[46]
[Link], [Link], A Machine Learning Approach For Plant Disease Using a Support vector machine (SVM) rice leaf disease like Bacterial leaf
[Link], Classification And Pesticides Suggestion Using blight, Brown spots& leaf Smut
[Link] (2020) Rank-Based Attribute Selection [13]
Multimedia Tools and Applications (2023) 82:36091–36117
Multimedia Tools and Applications (2023) 82:36091–36117 36101

& Due to inadequate knowledge or the limited availability of trained experts, it is challenging
for all growers to detect diseases.

Such problems exist in the previous work and need to be solved. So we need an expert model
to recognize the fast detection of diseases from rice leaves. To overcome such issues, we have
designed a model, namely a faster R-CNN with VGG-16 and a Random Forest classifier.

3.2 Methodology

The critical contribution of this technique is to use ML techniques to construct a rice disease
detection and classification model, which can be helpful for disease identification. UCI ML
Repository gathers the data for this venture. An open-source ML platform, Waikato Frame-
work for Knowledge Analysis (WEKA), has incorporated various ML algorithms to train the
system. But, our proposed simulation has been performed via MATLAB. This includes both
Disease identification and disease classification. In intelligent farming based on ML, the
device provides data on the crop employing RGB (red, green, blue) monitors, a standard
smartphone adapter, and other sources. Such an image is then processed offline and developed
by ML (machine learning) into valuable information – often without the additional expense of
multi-spectral monitoring and identifying the infection into crop image and classifying the type
of disease. There are various methods of Machine Learning.

3.2.1 Materials and methods

1. Image data description

2. Disease symptom
3. Disease Identification
4. Classification

Image data description The UCI ML (Machine Learning) Repository gathers the data for
this mission. The dataset has been manually separated into different illness groups by infected
leaves. We consulted and asked the farmers to send sample leaves with the names of the
diseases. We have defined and verified the English names of these conditions in consultation
with farm experts and have given names in their native languages (Gujarati). This dataset was
used for rice plant leaves diseases ıdentification & classification. As part of this work
following activities are conducted. (1) How different image features (2) can be obtained,
which image processing operations include the information required (3). The images were shot
in direct sunlight on a white backdrop. The pictures have been reduced to the required
processing resolution. The bacterial leaf bite, spot Brown and leaf smut, each having 40
images, are three classes/diseases. All pictures are in jpg format.

Disease symptom There are six types of Rice diseases, each with its symptoms (see Table 1).
On behalf of these symptoms, we can identify rice diseases.

Disease identification Here’s an easy solution to infection identification on rice crops that have
been applied to analyze the processes. The disease is identified as one of the following six
categories: Brown spots, Leaf smut, Bacterial leaf blight, sheath blight, Rice Blast, and Sheath rot.
36102 Multimedia Tools and Applications (2023) 82:36091–36117

Classification According to the definition of a plant disease, it is “anything that prohibits a

plant from achieving its greatest potential.” Plant Bacterial infections may be broadly catego-
rized as either contagious or non-infectious as per their primary causative agent. Living
pathogens or spores are responsible for bacterial crop plants. Such pathogens may transfer
to a healthy plant from an infected plant and wreckage; Nematodes, fungi, mycoplasmas, and
bacteria are microscopic organisms producing plant diseases.
Many machine learning models were employed to detect and identify pathogenic rice
bacteria. However, this research area seems to have tremendous potential for higher precision
in the ML (Deep Learning) subset developments. Multiple established/modified DL frame-
works and several visualization algorithms are employed to recognize and categorize the
causes of rice leaf diseases. The leaves were classified using the Faster R-CNN classification
with VGG-16 and RF (Random Forest) classification between 32 species.
The mathematical computation for the evolved proposed method is provided in this
subsection for each method, like Faster R-CNN classification, in which the loss function is
calculated and then uses the parameterizations of the four coordinates for calculating the
boundary regression. After this VGG-16 transfer learning model used layers wised calculation
and calculated the kernel filter and ReLU activation function in the convolution layer and
softmax function in the fully connected layer. Finally, the machine learning Random forest
classifier computes the Gini index and entropy to identify how nodes in a decision tree branch.

A. Faster R-CNN (Region-based Fully CNN) Deep Learning

There are two modules in the target detection system known as Faster R-CNN [40]. The first
module is a system that offers a deep, totally convolutional layer, and a Fast R-CNN sensor is
suggested as the second module for the regions. The whole framework is one extensive,
unified object recognition framework (as seen in Fig. 1). Utilizing the NNs (neural networks’)
recently prominent terminology of structures of ‘attention’, The module RPN specifies where
to look for the Fast R-CNN module.
A Region Proposal Network (RPNs) inputs any size of image & produces a collection of
proposals for rectangle-shaped structures, accompanying an abject performance. This section
is mentioned that the system constructs include such a completely convolution layers frame-
work. Since our key aim is to merge computation with a Faster R-CNN image system [24], we
also suppose that both networks have a typical collection of convolution layers. In different
Classifier

RoI Pooling

Proposals

Region Proposal Network

Feature maps

Conv layers

Image

Fig. 1 Faster R-CNN [21]

Multimedia Tools and Applications (2023) 82:36091–36117 36103

observations, we look at the Simonyan and Zisserman (VGG-16) model with 13 total
convolution layers to share.
We move a specific network around the last mutual convolution layer to produce selected
features over the convolution output feature production. Each small network uses a temporal
frame of the convolution function map data as an input. An increasing sliding window is
linked to a function of lesser dimensions (512-d for VGG, followed by ReLU). Such a function
is inserted into two completely connected sibling layers — a surface of package regression as
well as a layer of box classification. In this article, we included n = 3, considering that the
powerful receptive field upon this input image data is big (204 VGG).

Loss function To train RPNs, each connector is allocated a binary class mark (of becoming an
entity or not). They offer two kinds of anchors a positive mark: I anchor with a maximum
Intersection over Union (IoU) overlapped to a container of ground truth and (ii) anchored to an
IoU correlation with any ground-truth box greater than 0.7.
With all these concepts, we optimize an objective feature in faster R-CNN [24] regarding
multi-tasks failure. It determines the loss functions for a given image as follows in Eq. (1):
1 1
Lðfpig; ftigÞ ¼ ∑ Lcls ðpi; pi*Þ þ λ ∑ pi * Lreg ðti; ti*Þ ð1Þ
Ncls i Nreg i

Here, i termed the anchor index in a mini-batch & pi as the expected likelihood of anchor i
being an entity. Unless the anchor is positive or negative, the ground truth symbol is pi incl is 1
and 0. ti is a vector composed of the four parameterized coordinates of the predicted bounding
box. The ground-truth box is the bounding box coupled with the positive anchors. The lack of
classification Lcls is defined as loss of the log between 2 classes (object vs. not object). For RL
(regression loss), we choose Lreg (ti, ti∗) = R (ti – ti∗) wherever R is defined as a durable loss
feature (smoothen L1). The word pi ∗ Lreg indicates that the loss of regression is only enabled
for positive anchors (pi ∗ = 1) or otherwise disabled (pi ∗ = 0).The results of the class and
reg layers collectively contain {pi} and {ti}. The two categories are defined by Ncls and Nreg As
well as measured by a regulating factor λ (λ = 10).
In bounding box regression, we use the parameterizations of the four coordinates shown
below in Eq. (2) [25]:
tx ¼ ðx−xaÞ=wa
ty ¼ ðy−yaÞ=ha
tw ¼ logðw=waÞ
th ¼ logðh=haÞ
ð2Þ
tx* ¼ ðx*−xaÞ=wa
ty* ¼ ðy*−yaÞ=ha
tw* ¼ logðw*=waÞ
th* ¼ logðh*=haÞ

Wherein x, y, w, and h represent the box’s center coordinates, width, and height. The variables
x, xa, and x represent the anticipated box, anchor box, and ground truth box (for y, w, and h).
This is bounding-box regression from an anchor box to a neighboring ground-truth box.

Training RPNs The end-to-end RPN could be conditioned by back-propagation and SGD
(stochastic gradient descent). We adopt the testing technique “image-centric” to build the
model. The loss functions (LFs) for all anchoring nodes can be optimized, but it would skew
36104 Multimedia Tools and Applications (2023) 82:36091–36117

against the -ve class since they are governed. Usually, we select 256 anchors arbitrarily in an
image to determine a mini-batch LF (loss function). We automatically create all additional
layers with a standard deviation of 0.01 by taking the weights’ value from the”0-mean”
gaussian function. Almost all layers (mutual convolution layers) are configured using prepro-
cessing a framework to identify with ImageNet [41]. To preserve data, we balanced all layers
of conv3 1, including through the VGG net.

B. VGG-16 transfer learning approach

Firstly, With a CNN, the input images are processed via a series of layers, including
convolutional and pooling layers, flattening and FCLs. The CNN output is then created, which
classifies pictures. After creating CNN models from scratch, the model is fine-tuned using the
image augmentation approach. As a result, we use one of the pre-trained models – VGG-16 –
to categorize images & assess the accuracy of training & validation data.
When using DL, TL refers to using a recently trained network on a new challenge. TL is
frequently utilized in DL since it enables the network to be trained with a small amount of data
while retaining high accuracy and performance. TL occurs when a computer uses information
from one activity to increase generalization about someone else. The final few levels of the
trained network are changed with novel layers in TL, including an FCL and a softmax
classification layer with class labels of three in our work. We tried to open the layer in each
model and introduced a stack of one activation layer, one batch-normalization layer, and one
dropout layer. These models were evaluated with various dropout rates, learning rates, & batch
sizes. The input size used in VGG-16 is 224 × 224.
The following are the primary building components of a CNN:

Convolution Layer (CL) This layer’s feature map is constructed by running a kernel matrix
over the input matrix. A mathematical process is carried out known as convolution using the
slide of the Kernel matrix over the input array. The outcome of element-by-element matrix
multiplication is displayed on the feature map at every position. Using convolution is a
common practice in many fields, including image processing, statistical analysis, and physics.
More than one axis may be employed in convolution. If we have a 2D image input, I, as well
as a 2D kernel filter, K, we may compute the convoluted picture as follows in Eq. (3):
Sði; jÞ ¼ ∑m ∑n Iðm; nÞkði−m; j−nÞ ð3Þ

Non-linear activation functions (ReLU) In this node, we perform a nonlinear transformation

on the input signal after the convolutional layer has been applied. The ReLU is a piecewise
linear function that, if the input is positive, produces the same value as the input; otherwise, it
outputs 0. ReLUs is a recent innovation. The Eq. (4) is remarkably simplistic: max(0,z).

z z>0
RðzÞ ¼ ð4Þ
0 z≤0

Its derivative is achieved in Eq. (5)


0 1 z>0
R ðzÞ ¼ ð5Þ
0 z< 0
Multimedia Tools and Applications (2023) 82:36091–36117 36105

Pooling Layer (PL) The feature map output of the convolutional layer has the drawback of
storing the precise position of features in the input. This is the downside of this output. The
result will be a completely different feature map if the original image is cropped, rotated, or
subjected to minor modifications. Researchers suggest decreasing the sample size of the
convolutional layers as a solution to this problem. The ability to downsample can be achieved
by adding a pooling layer immediately following the nonlinearity layer. The representation
may generally be consistent with tiny input transformations by using pooling, which helps
with this. According to the principle of translation invariance, the results of the vast majority of
the pooling outcomes will not change unless the input is changed in some way, although only
slightly.

Fully Connected Layer The output of the final PL provides input to the FCL at the end of a
CNN. It is possible to have one or more of these levels present. Once the network is wholly
linked, this signifies that each node in the first layer is associated with each node in the second
layer. This softmax function has been used that is defined by the following Eq. (6):
  ezi
σ ! z ¼ K z ð6Þ
i ∑ j¼1 e j

wherein, σ is softmax, ! z is the input vector, ezi is a standard exponential function for input
vector, K is no. of classes in the multi-class classifier, ez j is a standard exponential function for
the output vector.

C. Random Forest

RF [43] is an ensemble of DT (decision trees) learning algorithms. Every other tree is equipped
with a randomly chosen N training dataset from a training collection of replacement N
samples. In this analysis, we have a set of {Ri = (xi, ci)} Ni = 1 training ROIs that are
collected arbitrarily to form each DT (decision tree). xi = {x1i, x2i, .. xpi} are p is set of image
extracting features within each ROI. The classification labels that depict past context or leaf
patches affected by multiple diseases are ci = {0, 1, .. C − 1}. The ensembles train each tree
as chooses to follow:

& The best split function for each tree node is chosen from a collection of features φ = {φk},
k = {0, 1, …K = 1},chosen randomly by the whole set of features. The set of
characteristics K in set φ is calculated as the quantity by which the system generates the
least OOB error rate, retaining a trade-off between the correlation between the trees and
each tree’s power.
& Each new test specimen xi is passed down through all the other N trees during the
classification process. The sample is assigned to a specific class {0, 1, .. C − 1} by each
tree. The specimen is implemented to the class estimated by most trees.

The Gini index, in Eq. (7) utilized to determine how nodes on a DT branch are connected, is
often utilized while creating RF, relying on classification data in our work.

Gini ¼ 1−∑Ci¼1 ðpi Þ2 ð7Þ

36106 Multimedia Tools and Applications (2023) 82:36091–36117

With this Eq. (7), the Gini of every branch on a node is calculated based on the class and
probability, allowing us to estimate which branches are more probable to appear. pi is the relative
frequency of the class in the dataset, whereas c denotes the number of classes being observed.
We may also use entropy (see in Eq. (8)) to identify how nodes in a DT branch in a different
way.

Entropy ¼ ∑Ci¼1 −pi * log2 ðpi Þ ð8Þ

It is possible to decide how the node should branch based on the probability of a specific result.
Because of the logarithmic function utilized in its calculation, it is more computationally
intensive than the Gini index.
Figure 2 shows the flowchart for the proposed Faster R-CNN with VGG-16 and RF model.
It visualizes the overall flow of rice leaves disease classification.

4 Results and discussion

This section provides experimentation to perform with the help of the MATLAB 2018
simulation tool. This section evaluates various experimental results. The UCI ML (Machine

Start

Images Rice Crop Images

database

Data preprocessing

Resize Enhancement Segmentation

Feature extraction

Split the data into train & test

Apply Random forest for classification

Train/Validate model

Disease Classification

No
Disease Further
Identified? Continue

Yes
Measure Performance

Stop

Fig. 2 Flowchart for proposed Faster R-CNN with VGG-16 and RF model
Multimedia Tools and Applications (2023) 82:36091–36117 36107

Learning) Repository gathers the data for this mission. The dataset has been manually divided
into various disease groups by infected leaves. This dataset was used for Rice Plant Disease
Identification. The images were reduced to the desired processing resolution. The bacterial leaf
bite, spot Brown and leaf smut, each having 40 images, are three classes/diseases. They are all
pictures in jpg format. The simulation results assess performance parameters such as accuracy,
precision, recall, specificity, and F-score.
Table 3 represents the Bacterial blight disease severity and their mean percentage for
different types of Genotypes for different ranges of years 2010 to 2011, 2011–2012, and
2012–2013.
Table 4 represents the mean value of 5 individual replicates from individual plants in
bacterial light at 0, 25, 50, and 100.
The line graph of means of 5 individual replicates for the given Table 4 in Bacterial blight
different values like 0, 25, 50, and 100 are represented in Fig. 3.
Figure 4 shows the segmentation and edge map process for identifying rice leaves disease.
İn this figure, brown spots identify as one of the diseases of rice leaves.
Figure 5 represents feature mapping after converting the image. İt shows the CNN layer
output after applying it to the rice leaves. İt consists of 5 layers of CNN to train and map the
rice leave image on the segmented image.
The impact of Cd on chlorophyll fluorescence (B) specifications in Bacterial blight leaves is
represented in Table 5. The measurements were performed on samples coming from individual
plants.

Table 3 Bacterial blight disease intensity of various Indian rice genotypes on leaves

Genotype Bacterial leaf blight disease severity (%)

2010–11 2011–12 2012–13 Mean

Varuna 27.1(31.3) 33.0(29.7) 28.8(32.5) 32.3(28.5)

DRMR-2800 12.9(21.0) 13.3(21.4) 14.1(22.0) 13.4(21.5)
DRMR-2801 12.8(21.1) 13.9(21.8) 13.1(21.2) 13.3(21.4)
DRMR-2802 12.3(20.5) 13.7(21.7) 13.7(21.7) 13.2(21.3)
DRMR-2803 10.9(19.3) 11.4(19.7) 12.0(20.2) 11.4(19.7)
DRMR-2804 16.2(23.7) 17.9(25.0) 16.5(24.0) 16.8(24.2)
DRMR-2805 9.4(17.9) 10.2(18.6) 11.5(19.8) 10.4(18.8)
DRMR-2806 8.3(16.7) 9.9(18.3) 10.7(19.1) 9.6(18.1)
DRMR-2807 13.4(21.5) 15.5(23.2) 16.7(24.1) 15.2(22.9)
DRMR-2808 14.6(22.5) 15.4(23.1) 14.3(22.2) 14.8(22.6)
DRMR-2809 15.4(23.1) 15.0(22.8) 15.8(23.4) 15.4(23.1)
DRMR-2810 16.3(23.8) 14.9(22.7) 16.2(23.7) 15.8(23.4)
DRMR-2811 15.8(23.4) 16.8(24.2) 16.3(23.8) 16.3(23.8)
DRMR-2812 17.1(24.4) 18.0(25.1) 18.1(25.1) 17.7(24.9)
DRMR-2813 17.2(24.5) 16.1(23.7) 17.1(24.4) 16.8(24.2)
EC-339000 15.5(23.2) 16.8(24.2) 15.8(23.4) 16.1(23.6)
EC-339296 16.8(24.2) 16.9(24.3) 25.0(17.8) 17.1(24.5)
EC-399313 14.9(22.7) 16.1(23.7) 14.9(22.7) 15.3(23.0)
EC-399301 13.4(21.5) 15.1(22.8) 15.7(23.3) 14.7(22.5)
EC-399299 13.3(21.4) 13.0(21.1) 13.4(21.4) 13.2(21.3)
PAB-9511 11.8(20.1) 13.3(21.4) 13.6(21.6) 12.9(21.0)
PAB-9534 12.4(20.6) 13.4(21.4) 14.1(22.1) 13.3(21.4)
PHR-2 12.6(20.7) 13.0(21.1) 13.9(21.9) 13.2(21.2)
LSD (P =0.05) 1.04 1.16 0.76 –

*Figures include parenthesis through angular translated quantities

36108 Multimedia Tools and Applications (2023) 82:36091–36117

Table 4 Means of 5 individual replicates coming from individual plants

Bj Pn Gs Ci E WUE

0 17.41 0.32 237.60 5.38 3.20

25 7.72 0.26 309.20 4.85 1.69
50 5.61 0.14 316.6 3.39 1.72
100 6.07 0.15 308.6 3.62 1.85

Pn: Variation of net CO2 assimilation,

Gs: Conductance of the stomata,
Ci: CO2 concentration intercellular,
E: perspiration,
WUE: water use efficiency under different Cd treatments, calculated as Pn/E.

The graphical representation of Cd′s impact on the specification of chlorophyll fluorescence(B)

in Bacterial blight leaves is represented in Fig. 6. The measurements were performed on samples
from individual plants in Bacterial blight with values like 0, 25, 50, and 100.
The graphical representation of the effect of Cd impact on photosystem activity between
Fv/fm and Qpsii in Bacterial blight leaves represents in Fig. 7. The maximum PSII quantum

350

300

250
0
200
25
150 50
100 100

50

0
Pn Gs Ci E WUE

Fig. 3 Means of 5 individual replicates

Brown marks indicate

spot of disease

Fig. 4 Segmentation and edge map for rice leave disease ıdentification
Multimedia Tools and Applications (2023) 82:36091–36117 36109

Fig. 5 Feature maps after an image conversion: (a) original image of rice image, (b) 1st convertible filter layer,
(c) 1st rectified o/p layer, (d) 2nd convolution filter layer, (e) 2nd o/p layer, (f) 3rd o/p layer, (g) 4th o/p layer, (h)
5th o/p layer

yield measurements and effective quantum efficiency of PSII were performed on samples in
Bacterial blight with different values like 0, 25, 50, and 100.
A distinction of suggested Faster R-CNN with VGG-16 through RF (Random-Forest)
classifier and hyperspectral images labeled with RGB, as shown in Fig. 8. The mark in red
shows the part contaminated with rust. This should be noted that Rust plots are probably
defined in the same way (see in the 1strow (ii) & (iii)), while a significant proportion of the
healthy object will be surrounded via a red mark in (ii) rather than (iii), that suggests a
misclassification using RF classifier.
Figure 9 represents the graphical visualization of testing accuracy for all images and
accurately predicted classes for those images. An x-axis indicates the number of classes &
the y-axis represents the total number of images. Each class shows how many % accurately
predicted images they have.

Table 5 Cd impact on photosystem activity

Bj Fo Fm Fv Fv/fm Fs Fm′ Qpsii

0 443.5 2285.0 1828.0 0.80 158.00 550.00 0.69

25 350.75 1328.0 996.0 0.75 208.50 607.25 0.64
50 366.00 1276.8 791.61 0.62 277.00 732.33 0.61
100 386.33 1366.3 888.09 0.65 211.00 651.40 0.62

F0: a minimum degree of fluorescence in an obscure state,

Fm: a maximal degree of fluorescence in the dark-adapted state,
Fv: fluoresce vector,
Fv / Fm: quantum yield of full PSII,
Fm: ‘the light-adapted specimens of total fluorescence,
Fs: fluorescence steady state,
Qpsıı: effective quantum efficiency of PSII.
36110 Multimedia Tools and Applications (2023) 82:36091–36117

Fm’

Fs
100
50
Fv
25
0
Fm

Fo

0 500 1000 1500 2000 2500

Fig. 6 Impact of Cd on specifications of Chlorophyll Fluorescence (B)

The x-axis of the graph in Fig. 10 indicates the network’s performance, via training
iterations, in the conducted test collection. The visualization of the model was acquired every
10,000 training processes. The blue line graph reflects the loss during the training phase. The
loss gradually decreased during the training process. The green line graph reflects the
validation accuracy during the validation phase. The accuracy has increased gradually de-
creased during the validation process.
Table 6 indicates the pathogenicity test of brown spots on rice leaves. İt has an inoculated
brown spot as a fungus but includes how much brown spot is under control or uninoculated. İt
comprises the number of inoculated leaves, the number of infected leaves, and how many % of
infected leaves. Each of divided into two categories uninjured and injured. This test shows that
91.30% of rice leaves are infected and injured due to brown spots, and 73.91% are infected and
uninjured due to brown spots.

100

50
Bj value

Fv/Fm
Qpsii
25

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

Fig. 7 Cd impacts photosystem activity between Fv/fm and Qpsii

Multimedia Tools and Applications (2023) 82:36091–36117 36111

Rust plot

Healthy plot

(i)Plot in RGB Color (ii) Rust Area (RF) (iii) Rust Area (Hybrid Model)
Fig. 8 UAV hyperspectral images: (i) a representation of RGB color, (ii) RF classifier, as well as (iii) developed
a hybrid system

4.1 Comparison and discussion

This subsection provides a comparison of the proposed method Faster R-CNN with VGG-16
and RF from previous works [40, 41, 43], including different machine learning and deep
learning algorithms, and this comparison is discussed in brief.
Table 7 provides a comprehensive comparison of all of the machine learning methods that
were investigated. According to the information presented in this table, several machine
learning (ML) algorithms were used to carry out a variety of experiments on rice leaves to
identify diseases that affect rice leaves. The evaluation of these methods was carried out with a
focus on the following five factors: precision, recall, accuracy, f-score, and specificity. The

Fig. 9 Accuracy is forecasted separately for each class

36112 Multimedia Tools and Applications (2023) 82:36091–36117

Fig. 10 Accuracy and loss of the proposed hybrid

proposed method is ideal for the research to assess which ML algorithms would produce the
most remarkable results, and we will pursue these techniques. It was discovered that the
proposed Faster R-CNN with VGG-16 and RF method obtained the highest performance
results among these algorithms. This is because it has a fast learning rate to obtain features of

Table 6 Pathogenicity test of brown spots on leaves of rice

Fungus inoculated Number of leaves inoculated Number of leaves infected % of leaves infected

Uninjured Injured Uninjured Injured Uninjured Injured

Brown spot 22 24 16 22 73.91% 91.30%

Control (Uninoculated) 50 50 – –

Table 7 Experimental comparative results for different ML techniques with the proposed method

Machine Learning Algo. Accuracy Precision Recall Specificity F-Score

KNN 88.67 78.14 90.92 80.72 84.05

NB 72.33 70.91 72.69 75 071.79
SSD with ResNet-50 80.23 82.53 96.33 83.89 79.16
MLR 80.24 24.17 96.66 0.0000 38.66
InceptionResNetV2 [38] 95.67 96.45 96.5 – 96.5
R-FCN [7] with ResNet-50 87.92 85.98 98.99 88.21 56.31
ANN 76.86 99.94 99.61 99.78 99.76
5-layer convolution model [51] 78.2 – – – –
Faster R-CNN with VGG-16 [40, 41] 88.64 83.06 92.09 79.87 76.98
Random Forest (RF) [43] 94.13 66.55 96.66 100 79.81
DENS-INCEP [16] 94.07 – 91.11 95.56 –
Faster R-CNN with VGG-16 and RF (Proposed) 97.3 99.94 99.64 99.96 99.91
Multimedia Tools and Applications (2023) 82:36091–36117 36113

Performance Measures

100
90
80
70
60
in %

50
40
30 Accuracy
20
10
0 Precision
Recall
Specificity

ML Algo

Fig. 11 Compares various machine learning algorithms to measure performance with the proposed hybrid model

images, and the random forest successfully classifies the images within the classes. It has an
accuracy rate of 97.3% for illness categorization, which is excellent in contrast to other
systems. Additionally, it has an accuracy of 99.94%, a recall of 99.64%, a specificity of
99.96%, and a value of 0.99 for its f-score.
In Fig. 11, we have compared different machine learning Algorithms to the proposed
method. From this Fig. 11, we can say that all performance measures for the proposed
algorithm perform outstandingly compared to other algorithms. Still, precision has the same
value in comparison to the ANN technique.
In Fig. 12, we have compared the F-score value for different machine learning techniques to
the proposed method. From this Fig. 12, we can say that the F-score value for the proposed
algorithm performs outstandingly compared to other algorithms.
From the result analysis, we have justified that the proposed faster R-CNN with VGG-16
and RF model has achieved an accurate classification rate compared to other state-of-art
methods. İt gives precise results to identify infected rive plant leaves injured or uninjured by
Brown spots on the leaves of rice. So overall, we can say that the proposed expert model
outperformed the accuracy, recall, specificity, precision, and F-score.

F-Score
1.5
1
0.5
0
F-Score

ML Algo

Fig. 12 Comparison of F-score value for various machine learning algorithms with the proposed hybrid model
36114 Multimedia Tools and Applications (2023) 82:36091–36117

5 Conclusion

Rice crops have been recognized as one of the most important energy sources for resource
development during the last few decades. Crop diseases pose a notable concern to sustenance
protection, but their rapid differentiation of proof remains uncomfortable worldwide due to the
absence of the essential base. The advent of precise techniques in classifying leaf-based images
has produced remarkable results. Although some DL designs/models have been established for
hyperspectral classification of images in the field of diagnosis the rice diseases, this is yet a
promising area of research and could therefore lead to different circumstances to help classify
rice diseases, such as different lighting conditions, considering the actual background, etc. The
early diagnosis of such a disease would protect the farmer from suffering a significant
economic loss. This proposed paper includes a hybrid DL (deep learning) and ML (machine
learning) techniques model (i.e., faster R-CNN including VGG-16 as well as RF) to recognize
the disease symptoms in the rice plant. Automatic crop disease detection is performed by using
implemented techniques. Disease in a plant results in decreased yields, which contributes to a
tremendous loss for farmers. Thus, early identification of disease would be helpful to farmers
such that appropriate steps can be taken to diminish the production of unhealthy crops. From
the experimental results, higher detection or classification rate of rice leaf disease has been
obtained in terms of accuracy, i.e., 97.3%, compared to other ML algorithms. Besides the
accuracy, the more performance parameters are recall, specificity, precision, and F-score, with
values of 99.64%, 99.96%, 99.94%, and 99.91%, respectively. From the analysis, it can be
said that all performance measures perform outstandingly compared to other algorithms, but
precision has the same values as the ANN technique.
The proposed approach can be combined in the future with several other methods yet to be
developed, like clustering techniques as well as NNS, disease detection techniques with
classification models, by using texture and color analysis to establish an expert framework
for the prediction and administering of early rice leaf disease. In the future, the system’s quality
can be improved by using advanced techniques for extracting the leaf entity from such a
complex context. Further needed to compute the amount of disease present on the leaf. This
work can be enhanced to other plant leaf diseases based on their properties, including early
warning systems, such as mustard, vegetables, cotton crops, fruits, beans, etc. Using other
additional techniques of deep learning like LSTM, RNN may be used to increase the system’s
performance for potential accuracy.

Data availability Data sharing not applicable to this article as no datasets were generated or analyzed during the
current study.

Declarations

Conflict of interest On behalf of all authors, the corresponding author states that there is no conflict of interest.

References

1. Abduljabbar R, Dia H, Liyanage S, Bagloee SA (2019) “Applications of artificial intelligence in transport:

An overview,” Sustain (Switzerland), [Link]
2. Ahmed K, Shahidi TR, Irfanul Alam SM, Momen S (2019) “Rice leaf disease detection using machine
learning techniques,”[Link]
Multimedia Tools and Applications (2023) 82:36091–36117 36115

3. Anjna, MS, Singh PK (2020) “Hybrid System for Detection and Classification of Plant Disease Using
Qualitative Texture Features Analysis,”[Link]
4. Arun Pandian J, Geetharamani G, Annette B (2019) “Data augmentation on plant leaf disease image dataset
using image manipulation and deep learning techniques,”[Link]
8971580
5. Baranwal S, Khandelwal S, Arora A (2019) “Deep learning convolutional neural network for apple leaves
disease detection,” SSRN Electron J, [Link]
6. Barbosa A, Trevisan R, Hovakimyan N, Martin NF (2020) “Modeling yield response to crop management
using convolutional neural networks,” Comput Electron Agric, [Link]
105197
7. Bari BS, Islam MN, Rashid M, Hasan MJ, Razman MAM, Musa RM, Ab Nasir AF, Abdul PP, Majeed A
(2021) A real-time approach of diagnosing rice leaf disease using deep learning-based faster R-CNN
framework. PeerJ Comput Sci 7:e432–e432. [Link]
8. Bhatti UA, Huang M, Wang H, Zhang Y, Mehmood A, Di W (2018) Recommendation system for
immunization coverage and monitoring. Hum Vaccin Immunother 14(1):165–171. [Link]
1080/21645515.2017.1379639
9. Bhatti UA, Huang M, Wu D, Zhang Y, Mehmood A, Han H (2019) Recommendation system using feature
extraction and pattern recognition in clinical care systems. Enterp Inf Syst 13(3):329–351. [Link]
1080/17517575.2018.1557256
10. Bhatti UA, Zeeshan Z, Nizamani MM, Bazai S, Yu Z, Yuan L (2022) Assessing the change of ambient air
quality patterns in Jiangsu Province of China pre-to post-COVID-19. Chemosphere 288:132569. [Link]
org/10.1016/[Link].2021.132569
11. Bhatti UA et al (2022) Local Similarity-Based Spatial&#x2013;Spectral Fusion Hyperspectral Image
Classification With Deep CNN and Gabor Filtering. IEEE Trans Geosci Remote Sens 60:1–15. https://
[Link]/10.1109/TGRS.2021.3090410
12. Bhatti UA, Yu Z, Hasnain A, Nawaz SA, Yuan L, Wen L, Bhatti MA (2022) Evaluating the impact of roads
on the diversity pattern and density of trees to improve the conservation of species. Environ Sci Pollut Res
29(10):14780–14790. [Link]
13. Bhavatarini CMT, Priyadharshini V, Radhakrishnan M (2020) A machine learning approach for plant
disease classification and pesticides suggestion using rank based attribute selection. Int J Sci Eng Res 11(3):
1021–1024
14. Bohra J, Sadhukhan VPR (2018) Management of brown spot disease in rice (Helminthosporium oryzae) by
spraying of cow urine. Int J Chem Stud 6(1):1721–1723
15. Chaudhari AK, Rakholiya KB, Baria TT (2019) “Epidemiological Study of False Smut of Rice (Oryza
sativa L.) in Gujarat,” Int J Curr Microbiol Appl Sci, [Link]
16. Chen J, Zhang D, Nanehkaran Y, Li D (2020) “Detection of rice plant diseases based on deep transfer
learning,” J Sci Food Agric, vol. 100, [Link]
17. Chokey T, Jain S (2019) “Quality assessment of crops using machine learning techniques,”[Link]
10.1109/AICAI.2019.8701294
18. Chopda J, Raveshiya H, Nakum S, Nakrani V (2018) “Cotton crop disease detection using decision tree
classifier,”[Link]
19. Cynthia ST, Shahrukh Hossain KM, Hasan MN, Asaduzzaman M, Das AK (2019) “Automated detection of
plant diseases using image processing and faster R-CNN algorithm,”[Link]
2019.9068092
20. de Bigirimana VP, Hua GKH, Nyamangyoku OI, Hòfte M (2015) “Rice sheath rot: An emerging ubiquitous
destructive disease complex,” Front Plant Sci, [Link]
21. “Faster R-CNN | ML,” (2020) geeksforgeeks
22. Ferentinos KP (2018) “Deep learning models for plant disease detection and diagnosis,” Comput Electron
Agric, [Link]
23. Ginanni K (2004) Peace Corps' information collection and exchange. Ser Rev 30(3):249–251
24. Girshick R (2015) “Fast R-CNN,”\, [Link]
25. Girshick R, Donahue J, Darrell T, Malik J (2013) “Rich feature hierarchies for accurate object detection and
semantic segmentation,” Proc IEEE Comput Soc Conf Comput Vis Pattern Recognit, [Link]
1109/CVPR.2014.81
26. Govardhan M, Veena MB (2019) “Diagnosis of tomato plant diseases using random Forest,”[Link]
10.1109/GCAT47503.2019.8978431
27. Hasan MJ, Mahbub S, Alom MS, Abu Nasim M (2019) “Rice disease identification and classification by
integrating support vector machine with deep convolutional neural network,”[Link]
ICASERT.2019.8934568
36116 Multimedia Tools and Applications (2023) 82:36091–36117

28. Johannes A et al (2017) “Automatic plant disease diagnosis using mobile capture devices, applied on a
wheat use case,” Comput Electron Agric [Link]
29. Kamal KC, Yin Z, Wu M, Wu Z (2019) “Depthwise separable convolution architectures for plant disease
classification,” Comput Electron Agric, [Link]
30. Kim WS, Lee DH, Kim YJ (2020) “Machine vision-based automatic disease symptom detection of onion
downy mildew,” Comput Electron Agric, [Link]
31. Kusumo BS, Heryana A, Mahendra O, Pardede HF (2019) “Machine learning-based for automatic detection
of Corn-Plant diseases using image processing,” [Link]
32. Lee SH, Goëau H, Bonnet P, Joly A (2020) “New perspectives on plant disease characterization based on
deep learning,” Comput Electron Agric, [Link]
33. Lu Y, Yi S, Zeng N, Liu Y, Zhang Y (2017) “Identification of rice diseases using deep convolutional neural
networks,” Neurocomputing, [Link]
34. Maniyath SR et al (2018) “Plant disease detection using machine learning,”[Link]
2018.00017
35. Nanjundan J et al (2020) “Identification of new source of resistance to powdery mildew of indian mustard
and studying its inheritance,” Plant Pathol J, [Link]
36. Naqvi SAH (2019) “Bacterial Leaf Blight of Rice: An Overview of Epidemiology and Management with
Special Reference to-Indian-Sub-Continent,” Pakistan J Agric Res, [Link]
2019/[Link]
37. Prajapati HB, Shah JP, Dabhi VK (2017) “Detection and classification of rice plant diseases,” Intell Decis
Technol, [Link]
38. Prasad KNLVN, Kumar CSP, Subedi B, Abraha HB, Sathishkumar VE (2021) Rice leaf diseases prediction
using deep neural networks with transfer learning. Environ Res 198:111275. [Link]
envres.2021.111275
39. Ramesh S, Vydeki D (2018) “Rice blast disease detection and classification using machine learning
algorithm,”[Link]
40. Ren S, He K, Girshick R, Sun J (2017) Faster R-CNN: towards real-time object detection with region
proposal networks. Adv Neural Inf Process Syst. [Link]
41. Russakovsky O et al (2015) “ImageNet large scale visual recognition challenge,” Int J Comput Vis, https://
[Link]/10.1007/s11263-015-0816-y
42. Saberi Anari M (2022) A hybrid model for leaf diseases classification based on the modified deep transfer
learning and ensemble approach for agricultural AIoT-based monitoring. Comput Intell Neurosci 2022:
6504616. [Link]
43. Sandika B, Avil S, Sanat S, Srinivasu P (2016) “Random forest based classification of diseases in grapes
from images captured in uncontrolled environments,”[Link]
44. Shahriar SA, Imtiaz AA, Hossain MB, Husna A, Eaty MNK (2020) “Review: Rice blast disease,” Annu Res
Rev Biol, [Link]
45. Sharif M, Khan MA, Iqbal Z, Azam MF, Lali MIU, Javed MY (2018) “Detection and classification of citrus
diseases in agriculture based on optimized weighted segmentation and feature selection,” Comput Electron
Agric [Link]
46. Sharma P, Hans P, Gupta SC (2020) “Classification of plant leaf diseases using machine learning and image
preprocessing techniques,”, [Link]
47. Shruthi U, Nagaveni V, Raghavendra BK (2019) “A review on machine learning classification techniques
for plant disease detection,”[Link]
48. Singh R, Singh GS (2017) “Traditional agriculture: a climate-smart approach for sustainable food produc-
tion,” Energy Ecol Environ, [Link]
49. Singh R, Sunder S, Kumar P (2016) Sheath blight of rice: current status and perspectives. Indian
Phytopathol 69(4):340–351
50. Singh AK, Sreenivasu SVN, Mahalaxmi USBK, Sharma H, Patil DD, Asenso E (2022) Hybrid feature-
based disease detection in plant leaf using convolutional neural network, Bayesian optimized SVM, and
random Forest classifier. J Food Qual 2022:2845320–2845316. [Link]
51. Sinha K, Ghoshal D, Bhunia N (2022) Rice leaf disease classification using transfer learning. Lect Notes
Networks Syst 375:467–475. [Link]
52. Tian K, Li J, Zeng J, Evans A, Zhang L (2019) “Segmentation of tomato leaf images based on adaptive clustering
number of K-means algorithm,” Comput Electron Agric, [Link]
53. Truong T, Dinh A, Wahid K (2017) “An IoT environmental data collection system for fungal detection in
crop fields,”[Link]
54. Verma G, Taluja C, Saxena AK (2019) “Vision based detection and classification of disease on Rice crops
using convolutional neural network,” [Link]
Multimedia Tools and Applications (2023) 82:36091–36117 36117

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a
publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such publishing agreement and
applicable law.