论文信息

Anshul Jaiswal / Naveen Kumar Dewangan / Amar Kumar Dey

International Journal of Electrical and Electronics Engineering

论文解读

AI 辅助整理
方法
文献综述,覆盖图像采集、预处理、特征表示、分类及评估指标全流程。
发现
研究范式已从手工特征机器学习转向深度学习,并亟需轻量模型与标准化基准。
机会点
可构建标准化跨域水稻病害基准数据集,并研究面向田间物联网的轻量化域自适应检测模型。

论文摘要

水稻叶片病害在发展中国家备受关注,尤其是在大量人口的生计依赖于农学的情况下。水稻被视为南亚国家的主要粮食作物,水稻叶面病原体造成严重的产量损失和经济负担。本综述系统考察了用于水稻叶片病害检测的机器学习及先进深度学习方法的发展演变。与仅关注模型性能的研究不同,本综述聚焦于完整的流程——涵盖图像采集过程中观察到的挑战、预处理步骤、数据表征、特征表示、所使用的分类方法,以及导致实际部署失败的误用评估指标。本文献揭示了一个明显的范式转变:从基于手工特征的机器学习系统转向深度学习架构,后者因其始终如一的高准确率、能够直接从原始图像中学习判别性表示而无需任何人工干预。本综述还强调了智慧物联网部署对轻量级模型日益增长的需求。总体证据表明,自动水稻叶片检测的未来取决于建立标准化基准数据集、稳健的跨域验证,以及与智慧农业框架相结合,融合认知计算机视觉和域自适应学习,从而能够在多样化的田间条件下实现准确的病害诊断。

全文

International Journal of Electrical and Electronics Engineering

SSRG International Journal of Electrical and Electronics Engineering Volume 13 Issue 10, 179-192, October 2026
ISSN: 2348-8379/ https://doi.org/10.14445/23488379/IJEEE-V13I10P113 © 2026 Seventh Sense Research Group®

This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Review Article
Review of Machine Learning Approaches for Leaf
Disease Detection in Rice Crop: Trends, Challenges, and
Future Opportunities
Anshul Jaiswal1, Naveen Kumar Dewangan2, Amar Kumar Dey3
1,2,3Department of Electronics and Telecommunication, BIT Durg, Chhattisgarh, India
3Corresponding Author: amardeyhope@gmail.com
Received: 24 June 2026 Revised: 01 October 2026 Accepted: 20 October 2026 Published: 28 October 2026
Abstract - Rice leaf disease is of important concern in developing countries, especially when the livelihood of a major population
is dependent on agronomy. Rice is considered a staple crop in South Asian countries, foliar pathogens in rice cause substantial
yield loss and economical burden. This review systematically examines the evolution of machine learning and state-of-the-art
deep learning approaches for the detection of rice leaf disea se. Unlike studies concentrated only on model performance, this
review focused on the complete pipeline - spanning challenges observed during image acquisition, preprocessing steps, data
characterization, features representation, classification methodology used, along with misused evaluation metrics leading to
unsuccessful real -world deployment. This literature reveals a clear paradigm shift from handcrafted feature -based machine
learning systems to deep learning architectures, because of their consistently high accuracy, learning discriminative
representations directly from raw images without any human intervention. This review also highlights the growing demand for
lightweight models for smarter IoT deployments. Overall evidence indicates that the future o f automatic rice leaf detection
depends on creating standardized benchmark datasets, robust cross-domain validation, and collaborating with smart agriculture
frameworks combined with cognitive computer vision and domain adaptive learning, which are capable of accurate disease
diagnosis under diverse field conditions.
Keywords - Deep learning, Handcrafted features, Leaf disease, Machine learning, Rice crop.

  1. Introduction
    Around fifty percent of the world population depends on
    Rice ( Oryza sativa ) for its caloric needs [1]. Specially in
    Southeast Asian countries, rice is considered the staple food
    and accounts for three -quarters of daily caloric intake [2].
    Ninety percent of global rice requirement is fulfilled by Asian
    countries, out of which Ind ia and China together contribute
    about 45% of the total Asian supply [3]. A majority of the
    population is directly or indirectly related to rice production
    for livelihood. Thus, the production of rice has a significant
    economic effect and food security con cern in developing
    countries.
    Leaf disease is considered one of the major constraints on
    rice productivity. Diseases like rice blast and bacterial blight
    make the leaf dry. On the other hand, sheath blight and brown
    spot reduce photosynthesis, leading to r estricted grain filling,
    and in severe cases, causing leaf death, resulting in substantial
    yield losses [4]. As reported [5], rice blast alone may reduce
    yield by 10 -30% and, if not controlled in time, severe
    outbreaks can wipe out entire fields. Similarly , a 60%
    production loss was reported by [6] due to bacterial blight.
    Pre-harvest crop loss accounts for 40% of global crop
    production, resulting in significant economic costs of about
    $40 billion annually.
    The conventional crop screening is slow, time -
    consuming, requires expertise, and is labour -intensive [7].
    Also, the disease symptoms appear late in the rice leaf ,
    resulting in delayed management. These challenges create a
    clear need for automated, scalable systems that can detect
    diseases before severe sy mptoms develop and reduce
    unnecessary chemical use. Therefore, early diagnosis offers
    one of the most effective ways to limit the crop from damage.
    Timely detection with accuracy optimizes pest and fungicide
    controls and reduces crop damage, resulting in h igh yield [8].
    Machine learning and deep learning algorithms have gained
    popularity in plant pathology over the past decade. Traditional
    models, like support vector machines and the random forests
    approach, rely on hand -crafted features extracted from leaf
    images, which are time -consuming and subject to limited
    feature extraction. On the other hand, Convolutional Neural
    Networks (CNNs) changed this scenario by learning spatial
    features directly from raw images. The majority of the studies

Amar Kumar Dey et al. / IJEEE, 13(10), 179-192, 2026

180
have achieved a high accuracy of 95% and above using CNN
algorithms, when trained with enough annotated samples, for
rice disease datasets [9]. Transfer learning and attention
mechanisms enhance the leveraging of knowledge and ability
to extract critical features when train ing data is limited. As a
result, machine learning and deep learning have become
necessary tools in crop leaf disease detection. The traditional
disease diagnosis involves visual inspection and human
intelligence. In large cultivation areas , conventional d isease
diagnosis is labour -intensive and time -consuming. Human
intervention in disease diagnosis increases the risk of errors,
specially in the early stage diagnosis or in the case of
overlapping disease symptoms. Variation of field conditions,
such as lig ht illumination (day light conditions) , background
clutter, and leave orientation , makes diagnosis difficult.
Geographical conditions like location and seasonal variations
further complicate the diagnosis process.
Machine learning and deep learning are evolving as a
quick and effective alternative to handle the above-mentioned
complexities. Many researchers reported model accuracy
above 95% using small and unbalanced data sets, evaluated
under controlled laboratory conditions. Performance of such a
laboratory-trained model severely degrades when subjected to
real and diverse field conditions. Therefore, a high model
accuracy is not a guarantee that it is ready for real field
application. Only Deep learning and machine learning
architectural reviews were comm only observed in existing
literature. Most of them were found reviewing algorithms or
comparing the accuracy values, whereas a quality rice leaf
disease diagnosis should include the combined influence of
image acquisition technique, process adopted, select ion of
diverse dataset, class imbalance effect, validation used, model
interpretability, computational requirement and most
importantly, field deployment capabilities.
Furthermore, while reporting the performance values, an
imbalanced dataset, non -equivalent classes and variation in
adopted evaluation metrics were frequently observed. Results
obtained by considering the above imbalance produce
misleading conclusions. Thus, in order to achieve a real field
deployment, it is not enough to only report which methods
give better performance. However, it also requires to be
documented why these improvements occur and weather this
improvement sustains with real field deployment. This review
addresses recent advances in deep learning and machine
learning techniques deployed for detecting rice leaf diseases.
It starts by focusing on the significance of the rice crop and
how foliar diseases impact crop yield, then extends to
imaging, followed by the preprocessing techniques required to
build reliable datasets. Three challenges , fist is the limited
dataset constraint, then inappropriate model interpretability
and finally, deployment issues, where observed during the
course of this survey, while comparing the performance of AI
and ML architectures for detecting th e rice leaf disease .
Finally, it outlines future directions for integrated smart
farming systems, including explainable AI. The goal of this
review is to clarify current trends and point to opportunities
for more robust disease diagnosis tools.
2. Image Ac quisition and Preprocessing
Techniques
All deep learning and ML algorithms rely on image
acquisition followed by preprocessing to extract significant
features for decision making. In order to achieve high
accuracy in real -world deployment, a critical revie w of
imaging devices used for the acquisition of leaf samples, along
with on -field acquisition challenges with possible
preprocessing approaches adopted to overcome the field
challenges, is discussed in the following sections.
2.1. Imaging Devices
Leaf phenotyping monitoring relies on a range of imaging
techniques, scaling from RGB to hyperspectral band s, each
suited to different complex tasks. Handheld devices ,
particularly smartphones , remain dominant due to their
imaging capabilities and computational diagnostic
applications [10, 11]. At larger scales, UAVs equipped with
hyperspectral, thermal, RGB, and multispectral sensors
provide high -resolution with significantly larger coverage
[12]. Which is already utilised for measuring leaf area index
along with chlorophyll estimation using hyperspectral band s
at nearly 400 to 1000 nanometre s [13]. The aerial viewing
ability of UAV makes it a promising technology for real-time
crop health monitoring ; at the same time , it is capable of
covering a wider area with fine spatial detail.
2.2. Lighting, Background, and Environmental Conditions
Image acquisition techniques are facing real -time
challenges like instability, propeller shadow and motion
blurriness, making the imag e acquisition task difficult. The
quality of data further degrade s due to random season
conditions, variation in daylight intensity and presence of
background clutter. It was found that lesion contrast decreases
rapidly under low light condition s, making e arly disease
detection even more challenging. Performance deviation in
terms of accuracy from 99% to 30% is observed, models
previously train on control data subjected to real -time data
[14].
2.3. Image Preprocessing Methods
To overcome the challenges of d ata acquisition , raw
images are subjected to three processing steps. Which
enhances the data quality and model performance. Some of the
popular preprocessing techniques are discussed below:
 Noise removal: To reduce sensor noise or blur, filters like
Gaussian, Median, Wiener, and Bilateral are commonly
employed. Motion blur ness issues are overcome using
filters such as Bilateral, Gaussian, Median and Wiener.
Image is further enhanced by the deconvolution or
masking [15].

Amar Kumar Dey et al. / IJEEE, 13(10), 179-192, 2026

181
 Segmentation: Real time images are o ften cluttered with
background noise. And segmentation is an important
preprocessing step that separates the reason of interest
from background clutter and further enhance s the
classification accuracy. Previously used colour
thresholding techniques are rep laced by Deep
segmentation models like UNet, Mask R -CNN and
DeepLab as found in recent studies [16].
 Normalization: Multispectral a nd hyperspectral systems
are often corrupted by variability between images. To
overcome it , spectral normalisation, reflectan ce
calibration and histogram equalization are common
choices among the researchers [17].
After combining the above techniques, models produce
cleaner and more consistent inputs, which further improves
the ML and DL models ’ dependability in field variabilit y or
uncontrolled settings. Prominent imaging devices, along with
the preprocessing approach adopted in recent studies , are
shown in Table 1.
Table 1. Summary of representative studies on image acquisition and preprocessing
Reference Imaging Device(s) Preprocessing Methods Notes / Outcomes
[18] UAV-mounted hyperspectral camera (400-1000 nm) PROSAIL inversion + Gaussian Process regression Accurately retrieved LAI and chlorophyll.
[19] Smartphone cameras Mobile capture; DL classification Citizen-science disease-diagnosis app.
[20] High-res RGB camera DeepLabv3+ segmentation ~0.77 IoU for wheat canopy under outdoor lighting.
[21] RGB field images Augmentation: brightness ±10-30%, noise Tested CNN robustness to varied conditions.
[22] UAV + smartphone Super pixel segmentation Developed a large auto-annotated dataset.
3. Datasets Used in Rice Leaf Disease Detection
The availability, diversity, and annotation quality of
image datasets are important factors that determine the
performance of machine learning models for the detection of
rice leaf disease. Over the past ten years, advancements in
dataset creation and processing have made it possible to create
detection frameworks that are more accurate and widely
applicable. In the following section, some important features
of custom and public datasets that directly affect model
performance are discussed.
3.1. Publicly Available Datasets
Publicly available sample leaf datasets play an important
role in comparing different algorithms and in repeated
measurements. These datasets contain high -resolution digital
images of healthy and infected leaf samples taken at different
geographical locations, with variable field and lighting
conditions. Some benchmark datasets mostly referred to by
researchers are discussed below.
 Rice Leaf Disease Detection (RLDD) Dataset: A very
high-resolution image dataset consisting of both diseased
and healthy leaf samples. The broader coverage and
annotated data make it a benchmark dataset for deep
learning and machine learning algorithms [23, 24].
 Rice Leaf Disease Image -Text Retrieval Dataset
(CRLDRD): A cross-modal dataset consisting of sample
images along with disease descriptions. Its purpose is to
improve the interpretability of real -world challenges in
leaf disease detection and to support multimodal retrieval
research [25, 26].
 PlantVillage and Related Collections: Used for training
single models to identify diseases of multiple plant
species simultaneously. Despite being a multivariate
dataset. It also includes valuable rice leaf disease
samples. That are extremely useful for transfer lear ning
experiments [27, 28].
 Existence of such a diverse dataset enhance s the
robustness and accuracy of disease classification models.
These datasets play a vital role in increasing the cross -
study comparability.
3.2. Dataset Characteristics: Size, Disease Classes, and
Format
A significant variation in the dataset is observed in most
of the recent studies. Some important features influencing the
model evaluation and learning problem complexity are
discussed below:
 Size: Open source dataset comprises 1000 to 10000
sample images. Most common publicly available datasets
like DeepMulti -Net ha ve a repository of 20000 full
images, and RLDD has around 4000 images [29, 30].
Custom data sets available usually contain less images but
are rich in contextual information, satisfying region -
specific acquisition protocols [31, 32].
 Disease Classes: Usually, rice leaf disease datasheets
have 3 to 9 classes, namely leaf blight, leaf blast, brown

Amar Kumar Dey et al. / IJEEE, 13(10), 179-192, 2026

182
spot, sheath blight and leaf smut [29, 33]. Few datasets
include further sub -classifications to highlight disease
severity stage and symptom morphology.
 Format: Majority of the dataset prefers image acquisition
in the RGB colour space under controlled or partially
controlled environment s. Recent efforts also include
multispectral images and paired image -text data for
multimodal learning [25, 34].
 Annotation: Disease categories are commonly labelled by
experts, but the level of detail and consistency can vary.
A small number of datasets include hierarchical or region-
level labels for detailed lesion segmentation and spatial
analysis [35, 36].
3.3. Custom and Real-Time Collected Datasets
Custom datasets overcome the constraints of public
corpora by encapsulating real -world variability in lighting,
leaf orientation, cultivar type, and di sease manifestation.
These datasets are usually collected in the field via handheld
cameras, UAVs, or IoT -enabled systems, making them more
ecologically valid [31, 37, 38]. Examples include: collections
of 3,000 to 6,000 field images organized into seven t o nine
disease classes [39 , 40]; DepMulti-Net’s 20,000-image field
dataset to help deep learning models work in real -world
situations [3 0]; and real -time IoT -based image streams for
constant monitoring and quick disease response [31]. Even
though these datasets show diversity in different regions, they
often lack the standardization needed for large -scale
benchmarking.
3.4. Dataset Challenges: Class Imbalance, Diversity, and
Annotation Quality
Despite achievements in data acquisition technology ,
challenges were identified that restrict data set us ability and
model generalization:
 Class Imbalance: Few pathogen classes, such as blast and
spot, are overrepresented, whereas disease s like tungro
and sheath blight are underrepresented [29, 33]. This
biased the model training process. To overcome the class
imbalance issue , augmentation and resampling are
applied.
 Limited Diversity: Custom data sets are region-based and
limited to regional and seasonal variations. This ha s
restricted cross-study applications [41, 38].
 Annotation Quality: Overlapping infection symptoms,
which are most prominent in the initial stages, may
sometimes confused subjective experts , which results in
label noise [25].
 Real-Time Deployment Issues: Models trained on
controlled datasets freque ntly fail under uncontrolled
field conditions where lighting fluctuations, cluttered
backgrounds, and motion blur reduce accuracy [31, 39].
Community-wide efforts in multi -institutional data
collection, dataset standardization, and enhanced annotation
procedures are required to overcome the above -mentioned
challenges.
Table 2. Summary of literature on datasets used in rice leaf disease detection
Study Dataset Type Size Disease Classes Format Source
Rice Leaf Disease Detection
Using ResNet50
Publicly
Available
High-res
images
Bacterial blight, Blast,
Brown spot, Tungro Images [29]
Rice Leaf Disease Detection
using Deep Learning
Publicly
Available 4,199 images 5 common rice leaf diseases Images [42]
Deep Learning-Based
Architectures for RLDD
Publicly
Available Large (varies) Multiple rice diseases Images [43]
Rice Paddy Disease Detection
Using CNN
Custom
Collected 5,932 images 9 disease types Images [44]
Automatic Detection in Various
Leaf Sizes
Publicly
Available 4,960 images 8 rice diseases Images [45]
Enhancing Real-time Detection
(VGG16)
Custom +
Benchmark
3,611 custom
images Multiple rice diseases Images [39]
DepMulti-Net (Large Dataset) Custom Collected 20,000 images 4 common rice diseases Images [30]
Alpha-EIOU-YOLOv8 Custom Collected 3,175 images Blast, Leaf Folder, Brown Spot Images [31]
CRLDRD Publicly Available Cross-modal 4 common rice leaf diseases Image and Text [25]
Rice Leaf Disease
Classification Fusion Vision
Custom
Collected 2,627 images Multiple rice leaf diseases Images [32]

Amar Kumar Dey et al. / IJEEE, 13(10), 179-192, 2026

183
4. Feature Extraction Techniques
Both traditional and deep learning-based rice leaf disease
detection systems rely heavily on feature extraction. Deep
neural representations that can learn hierarchical and task -
specific features have replaced manually created appearance
descriptors in this field over time.
The comparative efficacy of handcrafted, conventional,
and deep feature extraction techniques is reviewed in detail in
this section.
4.1. Handcrafted Features
4.1.1. Color Features
Leaf lesions are characterised by significant color
variations, captured by descriptors like colour moment (mean,
variance, skewness) and RGB histogram [33, 46].
These colour descriptors perform well in controlled or
laboratory conditions but are unreliable in real -time
applications exposed to natural lighting.
4.1.2. Texture Features
Texture analysis mainly incorporates measurements of
contrast, homogeneity, entropy, and the calculation of energy
from grey images. Texture features are good indicators of leaf
lesions and necrosis in plants [46]. From the above-mentioned
texture features, entropy and contrast are extensively used for
leaf disease detection.
4.1.3. Shape Features
Researchers make use of morphological details of lesions
extracted from disease contour geometry and shape
descriptors, such as disease area and hue moments. Hue
moments make the detection robust against rotation, scaling,
and translation.
However, effectiveness falls when lesion patterns are
partially covered due to clutter or having complex patterns
[46]. Combining color, shape, and te xture features together,
an accuracy of 94% is achieved [46]. Since this model depends
on handcrafted features, it is subject to field variations,
limiting performance in real field conditions [47].
4.2. Traditional Feature Descriptors
The following textur e-based feature descriptors provide
better robustness and are preferred over handcrafted
descriptors:
 Gray Level Co -occurrence Matrix (GLCM): Generates
characteristics like contrast, energy, correlation, and
homogeneity by capturing second -order spatial
relationships. When combined with SVM or Decision
Tree classifiers, GLCM -based descriptors frequently
perform better than more straightforward statistical
features [47].
 Histogram of Oriented Gradients (HOG): Encodes local
gradient orientation patterns for i dentifying edges and
lesion boundaries. HOG is beneficial for structural lesions
but less effective for diseases with diffuse or fine-grained
symptoms [48].
 Local Binary Patterns (LBP): Generates micro -texture
patterns with strong illumination invariance, but performs
poorly in cluttered backgrounds or overlapping infections
[47].
Comparative analyses consistently demonstrated that
while LBP and HOG show higher sensitivity to background
complexity, GLCM features achieved higher accuracy under
controlled imaging conditions.
4.3. Deep Feature Representations
Convolutional Neural Networks (CNNs): Deep CNNs
greatly outperform handcrafted and conventional descriptors
by enabling automatic hierarchical feature learning from raw
data.
CNNs perform well on large, w ell-annotated datasets by
capturing both low -level features and high -level semantic
attributes of disease symptoms [49, 50].
4.3.1. Transfer Learning
By fine -tuning on rice disease datasets, pre -trained
networks like ResNet, Inception, and VGGNet gained
widespread adoption. Numerous studies have shown that
transfer learning consistently increases accuracy and
decreases training time, with accuracies surpassing 95% [ 32,
48].
4.3.2. Attention Mechanisms and Hybrid Models
Attention modules are used i n more recent systems to
suppress irrelevant background noise and guide the model
toward pertinent lesion regions.
Attention-augmented DenseNets and CNN -LSTM
combinations further improve robustness in challenging
imaging scenarios [50, 51].
4.3.3. Performance Comparison and Trade-offs
Deep learning methods consistently outperform
conventional methods. A 97% accuracy rate was attained by
combining EfficientNet-B7 with HOG and Grad -CAM [48],
whereas attention -augmented CNNs outperformed baseline
architectures by 2-4% [51].
However, the use of deep models in real-time agricultural
systems is limited due to high computational and annotated
data requirements [30, 32].
A comparative study of feature extraction approaches is
presented in Table 3.

Amar Kumar Dey et al. / IJEEE, 13(10), 179-192, 2026

184
Table 3. Feature extraction technique comparison in recent peer-reviewed studies
Feature Extraction Method Algorithm Used Dataset [Reference] Accuracy (%)
RGB, Color Moments, Entropy, Shape SVM Public rice leaf dataset [46] 94
GLCM, LBP, HOG Decision Tree, k-NN Rice leaf images, custom [47] 92-94
HOG + Grad-CAM (hybrid) EfficientNet-B7 Rice Disease Dataset (2024) [48] 97
Deep CNN (VGG19 features) Light GBM (FVBC) Rice Leaf DB (2,627 images) [32] 97.6
CNN + Attention (LMN) ResNet + attention RPDD multi-class disease (2024) [50] 97.8
CNN + ROI Extraction (DenseNet) Hybrid Rice Disease field images [52] 96
Multi-Level Handcrafted Features Hybrid DL+ML Three benchmark datasets [33] Up to 99.5
4.4. Critical Insights and Outlook
Although crucial in early research, handcrafted and
conventional descriptors are constrained by their dependence
on controlled settings and manually designed features. Deep
neural representations now provide state -of-the-art accuracy,
especially when combined with attention and transfer learning
mechanisms. However, challenges remain in enabling
deployment on low -power devices and generalizing deep
models to highly variable field conditions. Current research
focuses on interpretable model architectures, domain
adaptation, and hybrid feature extraction techniques.
4.5. Machine Learning and Deep Learning Approaches
Automatic rice leaf detection mainly follows two
methodologies. In earlier works, handcrafted features were
extracted in the preprocessing stage and fed into conventional
machine learning classifiers. Recent trends highlight th e shift
of interest towards deep learning algorithms, mainly transfer
learning and CNN-based models, because leaf lesion features
are extracted from raw image datasets without human
intervention. Although the fundamental representation
learning of both ML and DL techniques is different, both
techniques focus on building a classification model that is easy
to implement, scalable, and accurate.
5. Machine Learning Approaches
The identification of rice leaf disease based on classical
machine learning technique s requires interdependent
preprocessing steps [53]. In this pipeline, the sample leaf
image is subjected to image normalization followed by
segmentation of the leaf lesion area for the extraction of
discriminative features. In the final step, dimensionalit y
reduction principles are applied to extract the most significant
features from the feature cluster [54, 55].
When classification is performed using Random Forest,
K-Nearest Neighbo ur (K NN), Support Vector Machines
(SVM), or Naive Bayes classifiers, the overall effectiveness of
the classifier is fundamentally dependent on the quality of
handcrafted feature extraction [56]. It is the visual quality of
features such as lesion patterns, morphological and statistical
signatures, extracted that determines the c lassification
accuracy.
5.1. Preprocessing, Segmentation, and Handcrafted
Descriptors
The digital image of a sample leaf acquired is generally
in RGB format. Due to variation in natural lighting conditions
and shadow, RGB data can sometimes mislead the disease
detection process. Therefore, the classical pipeline makes use
of color transformation like RGB to HSB or RGB to YCbCr
techniques, along with color histogram and morphological
operations, to segment the infected leaf area from the complex
background [57].
Threshold-based segmentation and clustering techniques
like K-means are extensively used to extract the disease area
from the surrounding healthy tissues [58 , 59]. Once
segmentation is implemented successfully, the following four
basic handcrafted descriptors are used to characterize the
infected leaf area: (1) color indices to capture chromatic
distortion and hue distribution of the necrotic tissue [60]; (2)
GLCM-derived texture features [61]; (3) structural descriptors
like profile contour and fractal measurements from disease
contour boundaries [62 , 63]; and (4) frequency -domain
features to discriminate microscopic details of similar disease
patterns [64, 65].
The classification performance increases by carefully
organizing the preprocessing. Special preprocessing
techniques like noise suppression, enhancement techniques
and elimination normalization increase feature stability. This
tree processing become s even more essential when samples
are collected from diverse feel conditions.
5.2. Traditional Classifiers and their Behaviours
From the literature, it is evident that Support Vector
Machine (SVM) is popular among researchers for its
consistently high performance because of its effectiveness
with small training sets in high -dimensional feature spaces
[66].
In situations when extracted feature sets are
heterogeneous, Random Forest and Gradient Boosting give
promising results [67]. K -Nearest Neighbour techniques are
distance-based algorithms effective in identifying spatial
relationships but suffer performance degradation with class

Amar Kumar Dey et al. / IJEEE, 13(10), 179-192, 2026

185
imbalances [67]. Lightweight classifiers such as Naive Bayes
are suitable for resource -constrained applications with a
compromise in classification accuracy.
Single-descriptor type datasets perform well in laboratory
conditions. Hyb rid datasets comprising colour, texture, and
shape descriptors outperform in the case of multi-evaluations.
These complex pipelines require domain expertise during
feature extraction. Any deviation in test conditions from
training conditions may deteriorate classification performance
[67, 68].
5.3. Advantages and Limitations of Classical Machine
Learning
Classical machine learning pipelines are particularly
advantageous in small-scale sample training applications and
offer faster computation [69]. Classical methods put the
foundation of machine learning intelligence and are
extensively used in lower-end devices and embedded systems.
However, classical pipelines face challenges in preprocessing
quality [70] and limited scalability [71], especially when
dealing with complex pathogen structure and multi -disease
datasets [72]. A detail comparative study of classical pipeline
focusing on its strengths and limitations is presented in Table
4.
Table 4. Comparative analysis of machine learning techniques for rice leaf disease detection (2013-2025)
Method Typical Accuracy Strengths Limitations Key Reference(s)
SVM (RBF, Linear) 94-97.5% Handles non-linear classes; robust to outliers Sensitive to kernel; struggles in high-dim. [33, 73, 74]
k-NN 80-90% Simple; effective for low/noise data Poor on large/heterogenous data; scalability [75]
Decision Tree 85-92% High interpretability Prone to overfitting, high variance [75]
Random Forest 95-98% Ensemble robustness; handles multi-class Reduced interpretability, parameter tuning [75, 76]
Ensembles
(Bagging/Boosting) 97-99% Superior accuracy; increased generalization Training time, complexity, and possible overfit [75, 77]
6. Deep Learning-based Approaches
The ability of deep learning algorithm s to extract the
special and textual features from raw input images makes it a
popular choice for the detection of rice leaf disease. The most
prefer deep learning architecture is the Convolution Neural
Network (CNN), because of thei r ability to identify local
disease patterns like lesion circumference, spot and chlorotic
area automatically [78].
The general procedure observed in implementing the DL
algorithm is, initially , a high -capacity model is selected ,
followed by training thro ugh inception or using transfer
learning. The model thus developed, found consistently out
perform in terms of accuracy, robustness as compared to the
traditional methods [24].
6.1. Convolutional Neural Networks and Transfer Learning
In DL architecture, th e upper layer of the Convolutional
Neural Network identifies irregularities in leaf colour due to
disease. Middle layer is engaged in exploring the structural
deformation that appears due to disease symptoms. Moreover,
the deepest layer identify the disease -specific morphology.
This multi -layer architecture facilitates rapid disease
deduction with an enhanced feature set.
Whereas in transfer learning, a high-performance model,
for example, VGG, AlexNet, GoogLeNet, ResNet, DenseNet,
and MobileNet [79, 80] is initialized first. The rice leaf disease
data set is used to fine -tune a pre-trained model. This type of
transfer learning architecture results in higher classification
performance even in the absence of larger data sets.
6.2. Modelling Imbalanced or Limited Data
One of the hurdle s in DL training is the limited
availability of rare disease datasheet s, which often creates a
cross-imbalance. DL overcome these hurdles by applying the
following techniques.:
 Data augmentat ion: In this approach , the data set is
artificially expanded, without introducing new images.
Bhai is generating transformation s like flipping and
rotation, and adjusting the brightness and contrast of the
image. T his alteration prevents the algorithm from
memorizing specific images.
 Class-balanced sampling: In this technique to approaches
are used for data visualisation. Presenting rare disease
samples more often, that is, over-sampling, or common
disease samples are shown lesser that is, under-sampling.
 Progressive training in this approach advances in steps .
First, it creates understanding by considering easy label
data, followed by learning a massive pool of unlabelled
or weekly labelled real-time data sets.

Amar Kumar Dey et al. / IJEEE, 13(10), 179-192, 2026

186
In the real world, leaf samples are cluttere d with
variations. To overcome this , GAN-based augmenters, along
with a style transfer network, are used, which create variations
in the data set, similar to field variation. Thus, make the
algorithm more robust to handle real-world challenges [81].
6.3. Lightweight and Edge-Deployable Architectures
Traditional deep learning models require higher and
processors like GPUs and TPUs, along with a continuous and
fast internet service. This makes disease detection in remote
locations challenging. Thus, in such sc enarios, lightweight
architectures like MobileNet, ShuffleNet, EfficientNet
(lighter version ) and SqueezeNet are effectively utilise d to
overcome hardware constraint s without much effecting the
classification performance [82] by offering a balance between
efficiency and accuracy, even for real-time field applications.
6.4. Beyond CNNs: Transformer Models and Hybrid
Designs
There is a Paradigm shift pain rice leaf detection from
classical CNN towards vision transformer and hybrid
architectures. When an image is subjected to a CNN, it process
the image using neighbourhood operation , and hence these
architectures are locally ba sed. Unlike the CNN transformer ,
it scans the entire leaf image and represents global context.
However, transformer architecture has a limited microscopic
view. The microscopic resolution capabilities of CNN
combined with the long -range dependencies of the
transformer are utilised in hybrid architectures. These
architectures are capable of extracting special features that
result in improved classifications when disease symptoms
have overlapping features.
6.5. Segmentation-Aware Deep Architectures
Another r ecent observation reviews a shift of method
from coarse image -level classification to find pixel level
segmentation. Traditional classifier s assign a single label to
the entire image, which makes multi -disease detection
challenging. Two overcome it ; a segmentation architecture is
introduced in this approach , and a specific class level is
assigned to individual pixels. This architecture is capable of
identifying healthy reason s for different pathogen symptoms
and background clitter for a given input image. I t was also
found that encode r-decoder architectures compress the
background letter efficiently, with promising early -stage
disease detection. [83, 84].
6.6. Strengths and Challenges of Deep Learning
This section highlights significant benefits and critica l
challenges faced during the deployment of a deep learning
architecture for rice leaves disease detection.
Strength of deep learning algorithm [85]:
 In traditional ML architecture , model performance
depends on the quality of handcrafted features extracted.
However, in DL architecture , features are extracted
automatically directly from the given input image pixels.
Which results in high performance as the abstract learning
of the fea tures by the DL layer provides a generalize
learning that enhances the unseen leaf features that
remain unexplored by the ML architecture.
 Real-time field variations remain inherent in traditional
ML architecture. DL take care of this issue by processing
the image at different layers, which provide spatial
hierarchies. So, DL architecture is more robust to field
diversity.
 A single leaf may have multiple stages, a level pathogen.
A single deep learning network can handle it easily by
updating its final classification layer. DL architectures are
scalable to multiple diseases.
 Early-stage disease detection is challenging as the disease
symptoms are visually indistinguishable by human
experts or traditional ML algorithm s. DL architecture
deep feature extractio ns with microscopic pixel -to-pixel
correlation and deep contextual relationship enhance its
ability to learn subtle visual clues.
Challenges in deep learning [86]:
 Millions of labelled data sets are required to prevent any
mathematical error.
 Required high er and graphical processor , making
standard on mobile deployment challenging.
 In case of a limited data set inquiry, the DL model may
overfit by collapsing into false predictions.
The DL offer significant prediction capabilities over
traditional ML models. It also offers some critical trade -offs,
such as overfitting when subjected to limited data set
conditions.
6.7. Integrated Perspective on ML and DL Approaches
ML and DL techniques complement each other in the
automatic detection of different stages and phases of rice leaf
disease. Classical pipelines suffer from limited feature
engineering dependencies but provide transparency and
efficiency. Deep learning requires higher -end hardware like
GPUs/TPUs to process large amounts of data and requires
huge labelled datasets since it starts learning from scratch.
The field rapidly integrates advanced deep learning
architectures such as CNN hybrids and segmentation -aware
models, which are lightweight. This technological update
gives a clear picture of the shift i n computer vision from
classical approaches to hybrid transfer CNN models and
segmentation-aware models, capable of identifying more
complex diseases using representation learning.
Performance analysis of prominent DL algorithms, found
in recent studies, along with the dataset used and deployment
feasibility, is provided in Table 5.

Amar Kumar Dey et al. / IJEEE, 13(10), 179-192, 2026

187
Table 5. Comparison of representative deep learning models
Model/Approach
[Reference] Dataset Accuracy (%) F1-Score (%) Key Features Deployment Feasibility
VGG19 Transfer Learning
[9]
Rice Leaf Image Set
(6 classes) 99 98 Fine-tuned, multi-class Mobile/UAV ready
MobileNetV2, ShuffleNet
[87]
Custom field
images 96-97 96 Lightweight, mobile-optimized Edge/mobile deployments
Improved Deep Residual
Networks [80]
Diverse rice disease
dataset 98.7 98.7 Attention mechanism added Medium—requires GPU/TPU
CNN+SVM Hybrid [52] Rice Leaf Disease Images 96 95 Feature selection, classic ML Highly adaptable, scalable
CNN-LSTM Pipeline [88,
89]
Time-series, video
data ~95 - Spatial-temporal analysis Experimental, field pilots
7. Performance Comparison and Evaluation
Metrics
A systematic approach was observed in ML and DL
models for the evaluation of rice leaf disease in the past
decade. More developed metrics, better reporting techniques,
and benchmark procedure s are in practice by researchers.
However, inconsistent datasets and variation in protocols used
make the comparison complicated.
7.1. Core Evaluation Metrics
Commonly used classification evaluation metrics
reported for leaf disease classification are AUC, F1-score,
recall, accuracy, and precision [90]. These metrics evaluate
how well a classifier can differentiate between a healthy and a
diseased sample. Accuracy, being intuitive, has a tendency to
hide imbalanced datasets, which is mostly common in disease
data.
This imbalance issue is well addressed by incorporating
precision and recall, which mainly focus on how a model
identifies positive cases and avoids false alarms. In studies
where both correct detection and error control matter, the F1-
score followed by the harmonic mean is preferred.
In the scenario where false positives and false negatives
have potential costs, AUC is preferred as it evaluates the
classification accuracy across the full range of thresholds [91].
It is very useful in early -stage disease detection. Some recent
literature also uses specificity, especially in large -scale
surveillance scenarios.
For segmentation, Intersection -over-Union (IoU) is
preferred for matching the segmented area with real -time leaf
segmentation.
It was found that the Matthews Correlation Coefficient
(MCC) was underutilized, as it can provide better summary
metrics when samples are collected from diverse fields.
7.2. Patterns in Cross-Study Comparisons
Variation in class distribution, variable daylight intensity,
and non-uniformity of sample datasets make it challenging to
compare different studies. However, a clear visible trend is
observed: deep learning models perform well over ML models
for both field -collected and laboratory datasets. Newly
developed models such as DepMulti -Net, DenseNet, and
MobileNetV2 report accuracy above 95%, further exceeded
above 98% using curated datasets [30].
Lighter architectures such as EfficientNet -B0,
MobileNetV2, and ShuffleNet are becoming popular for IoT
devices due to low power consumption, despite accuracy
constraints [87]. Object detection frameworks such as YOLO
and R -CNN-based models perform well in identifying and
classifying leaf lesions in real tim e, showing accuracy a bove
95% [23, 92].
This survey shows that when input samples are free from
background clutter and with controlled lighting conditions,
model performance is high. Performance decreases with the
introduction of field variation samples, sometimes very
significantly. There is a need for training CNN models with
field-variant datasets instead of curated datasets to handle real-
world challenges.
7.3. Dataset Characteristics and their Influence
Reported metrics are heavily influenced by datasets.
Laboratory datase ts reduce noise and enable even modest
architectures to achieve near-perfect accuracy. Weaknesses in
poorly regularized models are revealed by quality variability
introduced by publicly sourced datasets.
Field datasets are more trustworthy measures of rea l-
world performance. Leaf overlap, uneven illumination, and
temporally varying disease symptoms challenge models
assessed on such data. Studies that combine lab and field
images typically demonstrate better generalization.

Amar Kumar Dey et al. / IJEEE, 13(10), 179-192, 2026

188

Table 6. Summary of representative peer-reviewed research (past ten years)
Reference Model Dataset (Test zone) Evaluation Metrics Key Performance Outcomes
[30] DepMulti-Net Four rice diseases (lab) Accuracy 98.56% accuracy
[80] CBAM-based DRSN Three rice diseases (lab/field) mAP, Accuracy, Recall 98.89% mAP, 98.65% accuracy
[87] MobileNetV2, ShuffleNet Public & real-field images Accuracy, F1-score, AUC >96% accuracy, high F1/AUC
[79] ResNet197 Combined crop leaf dataset Accuracy 99.58% accuracy
[24] Faster R-CNN Online + in-field datasets Accuracy, Real-time performance >95% accuracy, real-time
[51] DenseNet, U-Net hybrid ROI-focused, lab images Accuracy 96% accuracy
[76] SVM, RF, KNN, VGG16 Laboratory images Accuracy, Precision, Recall, MCC >95% accuracy, diverse metrics
[47] DeepBatch Ensemble Field/lab mixed Accuracy, F1-score >97% accuracy, strong F1
7.4. Benchmark Models and Reproducibility Issues
Presently, the three models most widely adopted by
researchers as de facto benchmarks are: ResNet50 and its
higher variants for high classification accuracy demands;
ShuffleNet and MobileNetV2 for smartphone/IoT devices in
limited-resource scenarios; and S VM with manual feature
extraction as the classical ML baseline. These models are
well-documented, compatible with transfer learning, and
easily accessible. However, they often suffer from
reproducibility issues due to imbalanced public datasets and a
lack of proper documentation of experimental setup and
protocols. To achieve architectural development , there is a
need for standardize multi species open access data set. All
possible architectures must be evaluated using a common data
set for a coherent and fair validation. Recent models with key
performance features are tabulated in Table 6.
8. Conclusion
This survey extensively review s modern AI algorithm s,
which highlights the ultimate challenge s faced in the real -
world deployment. This survey critically reviews each step of
the image processing-based DL system from initial image
acquisition to final deployment of the intelligent model, with
an emphasis on future research. In the course of this survey, it
is discovered that automatic disease detection is a developing
research area and training the model in controlled laboratory
conditions presents serious deployment challenges in real field
conditions and ease of major concern s. This survey also t ook
into consideration the different image sensing modalities from
aerial to terrestrial, available for leaf disease detection. As the
image acquisition process directly affects the data set quality
and hence the model performance, the geographical, seasonal
and daylight intensity variation for the makes the data
acquisition more challenging. Lightweight classical ML
algorithms are still popular in smartphones, IoT and embedded
applications, but with a compromise in performance. In the
absence of hardware constraint s, DL architectures like CNN
and transfer learning -based algorithm s are the prime choice
among researchers because of their self-learning ability. A
clear shift of interest from handcrafted features to end -to-end
architectures is observed in the recent lea f disease
classification algorithm. It was found that some of the hybrid
algorithms and a few attention mechanisms perform
exceptionally better than conventional CNN models when
feed with a balanced and curated dataset. Small data sets and
a lack of standardized evaluation metrics mix the development
of auto matic rice leaf disease detection , challenging and
providing an I -score for future research. It was further
observed that some classifiers neglect failure evaluations; the
question of why the classifier fails in correct detection remains
unanswered. Classifier fal se positive prediction leads to the
agronomical resource wastage and false negative s, delay ing
the treatment process. These limitations make real -time field
implementation difficult. Now , in order to incorporate early
disease warning capabili ties, multispectral and hyperspectral
imaging outperformed RGB-based data acquisitions. The rice
leave detection is a reproductive research process , vision
based AI approach enable repeated disease monitoring which
again involves a need of open access stan dard data set, with
supportive policies framework for Cross study verification
and development of AI based rice leaf disease detection.
Overall, the available literature suggest s a potential but
incomplete deviation of experimental model s to operational
agricultural products . This requires a combined and aligned
effort of data scientist s, agronomist s and the Pharma
community. That will result in a cognitive AI -driven rice
disease detection system capable of making fair decisions
even in diverse field conditions.
Conflicts of Interest
No conflict of interest.
Funding Statement
No funding received.

Amar Kumar Dey et al. / IJEEE, 13(10), 179-192, 2026

189
References
[1] Jason P. Londo et al., “Phylogeography of Asian Wild Rice, Oryza Rufipogon, Reveals Multiple Independent Domestications of Cultivated
Rice, Oryza Sativa ,” Proceedings of the National Academy of Sciences , vol. 103, no. 25, pp. 9578 -9583, 2006. [CrossRef] [ Google
Scholar] [Publisher Link]
[2] Sumithra Muthayya et al., “An Overview of Global Rice Production, Supply, Trade, and Consumption,” Annals of the New York Academy
of Sciences, vol. 1324, no. 1, pp. 7-14, 2014. [CrossRef] [Google Scholar] [Publisher Link]
[3] Nirmala Bandumula, “Rice Production in Asia: Key to Global Food Security,” Proceedings of the National Academy of Sciences, India
Section B: Biological Sciences, vol. 88, no. 4, pp. 1323-1328, 2018. [CrossRef] [Google Scholar] [Publisher Link]
[4] Nur Akmal Rebitanim, Nur Zalikha Rebitanim, and Nur Shuhada Tajudin, “Impact of Silicon in Managing Important Rice Diseases: Blast,
Sheath Blight, Brown Spot and Grain Discoloration,” International Journal of Agronomy and Agricultural Research, vol. 6, no. 3, pp. 71-
85, 2015. [Google Scholar] [Publisher Link]
[5] Rohit Dhakarey et al., “Physiological and Proteomic Analysis of the Rice Mutant cpm2 Suggests a Negative Regulatory Role of Jasmonic
Acid in Drought Tolerance,” Frontiers in Plant Science, vol. 8, pp. 1-17, 2017. [CrossRef] [Google Scholar] [Publisher Link]
[6] P.R. Reddy, S.K. Nayak, and L. Bastiaans, “Simulation of the Effect of Bacterial Blight Disease on Crop Growth and Yield of Rice,”
Simulation and Systems Analysis for Rice Production (SARP), Wageningen, The Netherlands, Pudoc, pp. 340-347, 1991. [Google Scholar]
[Publisher Link]
[7] Ganeshamoorthy Hariharan, and Kandeeparoopan Prasannath, “Recent Advances in Molecular Diagnostics of Fungal Plant Pathogens: A
Mini Review,” Frontiers in Cellular and Infection Microbiology, vol. 10, pp. 1-14, 2021. [CrossRef] [Google Scholar] [Publisher Link]
[8] Ilaria Buja et al., “Advances in Plant Disease Detection and Monitoring: From Traditional Assays to In-Field Diagnostics,” Sensors, vol.
21, no. 6, pp. 1-22, 2021. [CrossRef] [Google Scholar] [Publisher Link]
[9] Ghazanfar Latif et al., “Deep Learning Utilization in Agriculture: Detection of Rice Plant Diseases using an Improved CNN Mod el,”
Plants, vol. 11, no. 17, pp. 1-17, 2022. [CrossRef] [Google Scholar] [Publisher Link]
[10] Xiaofei Yang et al., “Advances in UAV Remote Sensing for Monitoring Crop Water and Nutrient Status: Modeling Methods, Influencing
Factors, and Challenges,” Plants, vol. 14, no. 16, pp. 1-30, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[11] Aakash Chawade et al., “High-Throughput Field-Phenotyping Tools for Plant Breeding and Precision Agriculture,” Agronomy, vol. 9, no.
5, pp. 1-18, 2019. [CrossRef] [Google Scholar] [Publisher Link]
[12] M. Awais et al., “UAV -based Remote Sensing in Plant Stress Imagine using High -Resolution Thermal Sensor for Digital Agriculture
Practices: A meta -Review,” International Journal of Environmental Science and Technology , vol. 20, no. 1, pp. 1135 -1152, 2023.
[CrossRef] [Google Scholar] [Publisher Link]
[13] Kyu-Sung Lee et al., “Hyperspectral Versus Multispectral Data for Estimating Leaf Area Index in Four Different Biomes,” Remote Sensing
of Environment, vol. 91, no. 3-4, pp. 508-520, 2004. [CrossRef] [Google Scholar] [Publisher Link]
[14] Xiang Wang, Kai Wang, and Shiguo Lian, “A Survey on Face Data Augmentati on for the Training of Deep Neural Networks,” Neural
Computing and Applications, vol. 32, no. 19, pp. 15503-15531, 2020. [CrossRef] [Google Scholar] [Publisher Link]
[15] B. Vijayakumari, Noise Removal with Filtering Techniques, Advancements in Security and Privacy Initiatives for Multimedia Images, IGI
Global Scientific Publishing, pp. 133-156, 2021. [CrossRef] [Google Scholar] [Publisher Link]
[16] Archna Goyal, and Kamlesh Lakhwani, “A Du al-Pass Vision Transformer –U-Net Framework with Adaptive Feature Fusion for Citrus
Leaf Disease Segmentation and Classification,” Iran Journal of Computer Science, vol. 8, no. 4, pp. 2409-2430, 2025. [CrossRef] [Google
Scholar] [Publisher Link]
[17] Prasad S. Thenkabail, Remote Sensing Handbook, Volume I, 2nd ed., Boca Raton, CRC Press, 2024. [CrossRef] [Publisher Link]
[18] Rabi N. Sahoo et al., “Optimizing the Retrieval of Wheat Crop Traits from UAV -Borne Hyperspectral Image with Radiative T ransfer
Modelling using Gaussian Process Regression,” Remote Sensing, vol. 15, no. 23, pp. 1-18, 2023. [CrossRef] [Google Scholar] [Publisher
Link]
[19] Panagiotis Christakakis et a l., “Smartphone -based Citizen Science Tool for Plant Disease and Insect Pest Detection using Artificial
Intelligence,” Technologies, vol. 12, no. 7, pp. 1-21, 2024. [CrossRef] [Google Scholar] [Publisher Link]
[20] Radek Zenkl et al., “Outdoor Plant Segmentation with Deep Learning for High -Throughput Field Phenotyping on a Diverse Wheat
Dataset,” Frontiers in Plant Science, vol. 12, pp. 1-19, 2022. [CrossRef] [Google Scholar] [Publisher Link]
[21] Mohammad El Sakka et al., “A Review of CNN Applications in Smart Agriculture using Multimodal Data,” Sensors, vol. 25, no. 2, pp.
1-34, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[22] Sayali Shinde, and Vahida Attar, “An Indian Annotated Weed Dataset for Computer Vision Tasks in Precision Farming,” Data in Brief,
vol. 61, pp. 1-39, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[23] Kui Fang et al., “RLDD-YOLOv11n: Research on Rice Leaf Disease Detection based on YOLOv11,” Agronomy, vol. 15, no. 6, pp. 1-19,
2025. [CrossRef] [Google Scholar] [Publisher Link]
[24] Bifta Sama Bari et al., “A Real-Time Approach of Diagnosing Rice Leaf Disease using Deep Learning-based Faster R-CNN Framework,”
PeerJ Computer Science, vol. 7, pp. 1-27, 2021. [CrossRef] [Google Scholar] [Publisher Link]

Amar Kumar Dey et al. / IJEEE, 13(10), 179-192, 2026

190
[25] Hongliang Zhou et al., “A Precise Framework f or Rice Leaf Disease Image–Text Retrieval using FHTW-Net,” Plant Phenomics, vol. 6,
pp. 1-21, 2024. [CrossRef] [Google Scholar] [Publisher Link]
[26] Lei Chen et al., “Improved Domain Adaptive Rice Disease Image Recognition based on a Novel Attention Mechanism,” Computers and
Electronics in Agriculture, vol. 208, pp. 1-9, 2023. [CrossRef] [Google Scholar] [Publisher Link]
[27] SethuMadhavi Rallapalli, and M.A. Saleem Durai, “A Contemporary Approach for Disease Identification in Rice Leaf,” International
Journal of System Assurance Engineering and Management, 2021. [CrossRef] [Google Scholar] [Publisher Link]
[28] Gaurav Shrivastava, and Harish Patidar, “Rice Plant Disease Identification Decision Support Model using Machine Learning,” ICTACT
Journal on Soft Computing, vol. 12, no. 3, pp. 2619-2627, 2022. [CrossRef] [Google Scholar] [Publisher Link]
[29] D. Hemanth Kumar, and Samundiswary Punniakodi, “Rice Leaf Disease Detection using ResNet50,” 2024 International Conference on
Signal Processing, Computation, Electronics, Power and Telecommunication (IConSCEPT), Karaikal, India, pp. 1 -5, 2024. [CrossRef]
[Google Scholar] [Publisher Link]
[30] Kui Hu et al., “Identification of Rice Leaf Disease based on DepMulti-Net,” Frontiers in Plant Science, vol. 16, pp. 1-19, 2025. [CrossRef]
[Google Scholar] [Publisher Link]
[31] Dong Cong Trinh et al., “Alpha-EIOU-YOLOv8: An Improved Algorithm for Rice Leaf Disease Detection,” AgriEngineering, vol. 6, no.
1, pp. 302-317, 2024. [CrossRef] [Google Scholar] [Publisher Link]
[32] B. Naresh kumar, and S. Sakthivel, “Rice Leaf Disease Classification using a Fusion Vision Approach,” Scientific Reports, vol. 15, no. 1,
pp. 1-30, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[33] Yasmin M. Alsakar, Nehal A. Sakr, and Mohammed Elmogy, “An Enhanced Classification System of Various Rice Plant Diseases based
on Multi-Level Handcrafted Feature Extraction Technique,” Scientific Reports, vol. 14, no. 1, pp. 1-33, 2024. [CrossRef] [Google Scholar]
[Publisher Link]
[34] Jiaqi Li et al., “An Interpretable High -Accuracy Method for Rice Disease Detection based on Multisource Data and Transfer Learning,”
Plants, vol. 12, no. 18, pp. 1-22, 2023. [CrossRef] [Google Scholar] [Publisher Link]
[35] Tianqi Wei et al., “A Large -Scale In-the-wild Dataset for Plant Disease Segmentation,” Scientific Data, vol. 13, no. 1, pp. 1 -15, 2026.
[CrossRef] [Google Scholar] [Publisher Link]
[36] Qiaoyi Xu et al., “A Review of Advanced Deep Learning Methods of Multi-Target Segmentation for Breast Cancer WSIs,” IEEE Access,
vol. 13, pp. 76016-76037, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[37] Afif Hasane et al., “A Comprehensive Dataset of Rice Leaf Images for Disease Detection using Machine Learning,” Data in Brief, vol.
62, pp. 1-14, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[38] Chinna Gopi Simhadri, and Hari Kishan Kondaveeti, “Automatic Recognition of Rice Leaf Diseases using Transfer Learning,” Agronomy,
vol. 13, no. 4, pp. 1-24, 2023. [CrossRef] [Google Scholar] [Publisher Link]
[39] Saba un Nisa, Khalid Hamza, and Abdullah Shafiq, “Enhancing Real-Time Detection of Rice Diseases using an Optimized Deep Learning
Model,” Plant Protection, vol. 8, no. 4, pp. 679-689, 2024. [CrossRef] [Google Scholar] [Publisher Link]
[40] Hua Yang et al., “Disease Detection and Identification of Rice Leaf based on Improved Detection Transformer,” Agriculture, vol. 13, no.
7, pp. 1-17, 2023. [CrossRef] [Google Scholar] [Publisher Link]
[41] A. Bala Ayyappan et al., “Rice Plant Disease Detection using Convolutional Neural Networks,” Discover Artificial Intelligence, vol. 5,
no. 1, pp. 1-10, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[42] Ravindra Gahane et al., “Rice Leaf Disease Detection using Deep Learning,” International Journal for Research in Applied Science and
Engineering Technology, vol. 12, no. 4, pp. 1191-1193, 2024. [Google Scholar] [Publisher Link]
[43] Prameetha Pai et al., “Deep Learning-based Automatic Diagnosis of Rice Leaf Diseases using Ensemble CNN Models,” Scientific Reports,
vol. 15, no. 1, pp. 1-13, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[44] Jagamohan Padhi et al., “Enhanci ng Paddy Leaf Disease Diagnosis - A Hybrid CNN Model using Simulated Thermal Imaging,” Smart
Agricultural Technology, vol. 10, pp. 1-15, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[45] Kantip Kiratiratanapruk et al., “Automatic Detection of Rice Disease in Images of Various Leaf Sizes,” IET Image Processing, vol. 19,
no. 1, pp. 1-14, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[46] Anusha Preetham et al., “Instinctive Recognition of Pathogens in Rice using Reformed Fractional Differential Segmentation and
Innovative Fuzzy Logic -based Probabilistic Neural Network,” Journal of Food Quality , vol. 2022, no. 1, pp. 1 -17, 2022. [CrossRef]
[Google Scholar] [Publisher Link]
[47] Md. Mehedi Hasan et al., “Machine Learning and Image Processing Techniques for Rice Disease Dete ction: A Critical Analysis,”
International Journal of Plant Biology, vol. 14, no. 4, pp. 1190-1207, 2023. [CrossRef] [Google Scholar] [Publisher Link]
[48] Md. Shohanur Islam Sobuj et al., “Leveraging Pre -trained CNNs for Efficient Feature Extraction in Rice Leaf Disease Classification,”
2024 International Conference on Advances in Computing, Communication, Electrical, and Smart Systems, Dhaka, Bangladesh, pp. 1-6,
2024. [CrossRef] [Google Scholar] [Publisher Link]
[49] Verry Riyanto et al., “Deep Learning Approaches for Plant Disease Diagnosis Systems: A Review and Future Research Agendas,” Journal
of Applied Agricultural Science and Technology, vol. 9, no. 2, pp. 185-204, 2025. [CrossRef] [Google Scholar] [Publisher Link]

Amar Kumar Dey et al. / IJEEE, 13(10), 179-192, 2026

191
[50] Pengtao Lv et al., “An Improved Multi -Scale Feature Extraction Network for Rice Disease and Pest Recognition,” Insects, vol. 15, no.
11, pp. 1-25, 2024. [CrossRef] [Google Scholar] [Publisher Link]
[51] Mayuri Sharma et al., “Identification of Rice Leaf Diseases and Deficiency Disorders using a Novel DeepBatch Technique,” Open Life
Sciences, vol. 18, no. 1, pp. 1-21, 2023. [CrossRef] [Google Scholar] [Publisher Link]
[52] Wufeng Liu, Liang Yu, and Jiaxin Luo, “A Hybrid Attention -Enhanced DenseNet Neural Network Model based on Improved U-Net for
Rice Leaf Disease Identification,” Frontiers in Plant Science, vol. 13, pp. 1-16, 2022. [CrossRef] [Google Scholar] [Publisher Link]
[53] Prabira Kumar Sethy et al., “Deep Feature based Rice Leaf Disease Identification using Support Vector Machine,” Computers and
Electronics in Agriculture, vol. 175, pp. 1-9, 2020. [CrossRef] [Google Scholar] [Publisher Link]
[54] Itziar Fernández, Andrea Novo-Diez, and María J. González-García, “A Modular and Adaptable Approach for Automated Morphological
Feature Extraction in Meibography Images,” Scientific Reports, vol. 15, no. 1, pp. 1 -17, 2025. [CrossRef] [Google Scholar] [Publisher
Link]
[55] Habiba Njeri Ngugi, Andronicus A. Akinyelu, and Absalom E. Ezugwu, “Machine Learning and Deep Learning for Crop Disease
Diagnosis: Performance Analysis and Review,” Agronomy, vol. 14, no. 12, pp. 1-39, 2024. [CrossRef] [Google Scholar] [Publisher Link]
[56] Phan Thanh Noi, and Martin Kappas, “Comparison of Random Forest, k -Nearest Neighbor, and Support Vector Machine Classifiers for
Land Cover Classification Using Sentinel -2 Imagery,” Sensors, vol. 18, no. 1, pp. 1 -20, 2018. [CrossRef] [Google Scholar] [Publisher
Link]
[57] Snehaprava Acharya et al., “Automated Detection of Rice Plant Diseases using Dual Stage Thresholding and Twin Support Vector
Machine,” Bulletin of Electrical Engineering and Informatics, vol. 15, no. 1, pp. 512-521, 2026. [CrossRef] [Google Scholar] [Publisher
Link]
[58] Hassan Muhammad Yusuf et al., “A Systematic Review of Deep Learning Techniques for Rice Disease Recognition: Current Trends and
Future Directions,” Franklin Open, vol. 8, pp. 1-14, 2024. [CrossRef] [Google Scholar] [Publisher Link]
[59] Shuo Chen et al., “An Approach for Rice Bacterial Leaf Streak Disease Segmentation and Disease Severity Estimation,” Agriculture, vol.
11, no. 5, pp. 1-18, 2021. [CrossRef] [Google Scholar] [Publisher Link]
[60] S. Abisha, and T. Jayasree, “Application of Image Processing Techniques and Artificial Neural Network for Detection of Diseas es on
Brinjal Leaf,” IETE Journal of Research, vol. 68, no. 3, pp. 2246-2258, 2022. [CrossRef] [Google Scholar] [Publisher Link]
[61] R.A. Saputra et al., “Rice Leaf Disease Image Classifications using KNN based on GLCM Feature Extraction,” Journal of Physics:
Conference Series, vol. 1641, pp. 1-7, 2020. [CrossRef] [Google Scholar] [Publisher Link]
[62] Yijun Hu, Jingfang Shen, and Yonghao Qi, “Estimation of Rice Biomass at Different Growth Stages by using Fractal Dimension in Image
Processing,” Applied Sciences, vol. 11, no. 15, pp. 1-18, 2021. [CrossRef] [Google Scholar] [Publisher Link]
[63] Vishnu Muraleedharan, Sajeev C. Rajan, and R. Jaishanker, “Determining the Limits of Traditional Box -Counting Fractal Analysis in
Leaf Complexity Studies,” Flora, vol. 304, pp. 1-8, 2023. [CrossRef] [Google Scholar] [Publisher Link]
[64] Dongxue Zhao et al., “Study on the Classification Method of Rice Leaf Blast Levels based on Fusion Features and Adaptive -Weight
Immune Particle Swarm Optimization Extreme Learning Machine Algorithm,” Frontiers in Plant Science , vol. 13, pp. 1 -15, 2022.
[CrossRef] [Google Scholar] [Publisher Link]
[65] J. Dhakshayani, and B. Surendiran, “GF-CNN: An Enhanced Deep Learning Model with Gabor Filters for Maize Disease Classification,”
SN Computer Science, vol. 4, no. 5, pp. 1-12, 2023. [CrossRef] [Google Scholar] [Publisher Link]
[66] Amit Bhola, and Prabhat Kumar, “Deep Feature-Support Vector Machine based Hybrid Model for Multi-Crop Leaf Disease Identification
in Corn, Rice, and Wheat,” Multimedia Tools and Applications , vol. 84, no. 8, pp. 4751 -4771, 2024. [CrossRef] [ Google Scholar ]
[Publisher Link]
[67] Silky Sachar, and Anuj Kumar, “Survey of Feature Extraction and Classification Techniques to Identify Plant through Leaves,” Expert
Systems with Applications, vol. 167, pp. 1-14, 2021. [CrossRef] [Google Scholar] [Publisher Link]
[68] Chengzhuan Yang, “Plant Leaf Recognition by Integrating Shape and Texture Features,” Pattern Recognition, vol. 112, pp. 1-15, 2021.
[CrossRef] [Google Scholar] [Publisher Link]
[69] Gonzalo Miranda Cabrera, and Clemente Rubio-Manzano, “Image Classification using Deep and Classical Machine Learning Models on
Small Datase ts: A Complete Comparative,” CLEI Electronic Journal , vol. 27, no. 1, pp. 1 -35, 2024. [CrossRef] [ Google Scholar ]
[Publisher Link]
[70] Tala Talaei Khoei, and Naima Kaabouch, “Machine Learning: Mod els, Challenges, and Research Directions,” Future Internet, vol. 15,
no. 10, pp. 1-29, 2023. [CrossRef] [Google Scholar] [Publisher Link]
[71] Pegah Mavaie, Lawrence Holder, and Michael K. Skinner, “Hybrid Deep Learning Approach to Improve Classification of Low -Volume
High-Dimensional Data,” BMC Bioinformatics, vol. 24, no. 1, pp. 1-20, 2023. [CrossRef] [Google Scholar] [Publisher Link]
[72] Maarten R. Dobbelaere et al., “Machine Learning in Chemical Engineering: Strengths, Weaknesses, Opportunities, and Threats,”
Engineering, vol. 7, no. 9, pp. 1201-1211, 2021. [CrossRef] [Google Scholar] [Publisher Link]
[73] Lei Feng et al., “Investigation on Data Fusion of Multisource Spectral Data for Rice Leaf Diseases Identification using Machine Learning
Methods,” Frontiers in Plant Science, vol. 11, pp. 1-17, 2020. [CrossRef] [Google Scholar] [Publisher Link]

Amar Kumar Dey et al. / IJEEE, 13(10), 179-192, 2026

192
[74] Komal Bashir, Mariam Rehman, and Mehwish Bari, “Detection and Classification of Rice Diseases: an Automated Approach using
Textural Features,” Mehran University Research Journal of Engineering and Technology , vol. 38, no. 1, pp. 239 -250, 2019. [Google
Scholar] [Publisher Link]
[75] Majji V Applalanaidu, and G. Kumaravelan, “A Review of Machine Learning Approaches in Plant Leaf Disease Detection and
Classification,” 2021 Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks ,
Tirunelveli, India, pp. 716-724, 2021. [CrossRef] [Google Scholar] [Publisher Link]
[76] Ruyue Li et al., “Predicting Rice Diseases using Advanced Technologies at Different Scales: Present Status and Future Perspec tives,”
aBIOTECH, vol. 4, no. 4, pp. 359-371, 2023. [CrossRef] [Google Scholar] [Publisher Link]
[77] Le Yang et al., “Stacking -based and Improved Convolutional Neural Network: A New Approach in Rice Leaf Disease Identification,”
Frontiers in Plant Science, vol. 14, pp. 1-15, 2023. [CrossRef] [Google Scholar] [Publisher Link]
[78] Burak Gülmez, “RETRACTED: Advancements in Rice Disease Detection through Convolutional Neural Networks: A Comprehensive
Review,” Heliyon, vol. 10, no. 12, pp. 1-16, 2024. [CrossRef] [Google Scholar] [Publisher Link]
[79] J. Arun Pandian et al., “An Improved Deep Residual Convolutional Neural Network for Plant Leaf Disease Detection,” Computational
Intelligence and Neuroscience, vol. 2022, pp. 1-9, 2022. [CrossRef] [Google Scholar] [Publisher Link]
[80] Yang Lu et al., “New Method for Rice Disease Identification based on Improved Deep Residual Shrinkage Network,” Systems Science
and Control Engineering, vol. 11, no. 1, pp. 1-10, 2023. [CrossRef] [Google Scholar] [Publisher Link]
[81] Syed Taha Yeasin Ramadan et al., “Generative Adversarial Network -based Augmented Rice Leaf Disease Detection using Deep
Learning,” 2022 25th International Conference on Computer and Information Technology , Cox's Bazar, Bangladesh, pp. 976-981, 2022.
[CrossRef] [Google Scholar] [Publisher Link]
[82] Amit Kumar Manjhvar, and Rajendra Parmula, “A Lightw eight Deep Learning Architecture for Potato Leaf Disease Detection: A
Comprehensive Survey,” bioRxiv, pp. 1-20, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[83] Neha Sharma et al., “DBA-DeepLab: Dual-Backbone Attention-Enhanced DeepLab V3+ Model for Plant Disease Segmentation,” Food
Science and Nutrition, vol. 13, no. 7, pp. 1-28, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[84] Jianian Li et al., “Lightweight Rice Leaf Spot Segmentation Model based on Improved Deeplabv3+,” Frontiers in Plant Science, vol. 16,
pp. 1-19, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[85] Kazi Naimur Rahman et al., “A Real Time Monitoring System for Accurate Plant Leaves Disease Detection using Deep Learning,” Crop
Design, vol. 4, no. 1, pp. 1-16, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[86] Thandiwe Nyawose, Rito Clifford Maswanganyi, and Philani Khumalo, “A Review on the Detection of Plant Disease using Machine
Learning and Deep Learning Approaches,” Journal of Imaging, vol. 11, no. 10, pp. 1 -37, 2025. [CrossRef] [Google Scholar] [Publisher
Link]
[87] Vishan Kumar Gupta et al., “Classification of Rice Disease using Lightweight CNN Architectures,” 2026 6th International Conference on
Intelligent Technologies, Belgavi, India, pp. 1-6, 2026. [CrossRef] [Google Scholar] [Publisher Link]
[88] Poorva Vishnoi et al., “Automated Disease Detection of Rice Crops through Video Recognition,” 2025 3rd International Conference on
Device Intelligence, Computing and Communication Technolog ies, Dehradun, India, pp. 694 -697, 2025. [CrossRef] [Google Scholar]
[Publisher Link]
[89] Pardeep Seelwal et al., “A Systematic Review of Deep Learning Applications for Rice Disease Diagnosis: Current Trends and Fut ure
Directions,” Frontiers in Computer Science, vol. 6, pp. 1-20, 2024. [CrossRef] [Google Scholar] [Publisher Link]
[90] Usman Idris Ismail et al., “A Comprehensive Review of Deep Learning Approaches for Rice Disease Detection: Datasets, Methodologies,
and Future Directions,” Smart Agricultural Technology, vol. 11, pp. 1-11, 2025. [CrossRef] [Google Scholar] [Publisher Link]
[91] Sheethal Bhat et al., “AUCReshaping: Improved Sensitivity at High -Specificity,” Scientific Reports , vol. 13, no. 1, pp. 1 -16, 2023.
[CrossRef] [Google Scholar] [Publisher Link]
[92] Canlin Pan et al., “SSD-YOLO: A Lightweight Network for Rice Leaf Disease Detection,” Frontiers in Plant Science, vol. 16, pp. 1-16,
2025. [CrossRef] [Google Scholar] [Publisher Link]

论文原文与出版信息

International Journal of Electrical and Electronics Engineering
查看论文原文

继续探索