论文信息

Quart-ul-ain Abro / Junjie Xiang / Wazir Muhammad 等 6 人

Frontiers in Plant Science

论文解读

AI 辅助整理
方法
用SAR-CLD-2024数据集9137张图像,迁移学习对比十种CNN框架。
发现
EfficientNetB0平均准确率最高达91.29%,VGG19最低仅57.63%。
机会点
可探索轻量化模型在边缘设备上的实时病害诊断与不平衡数据鲁棒性优化。

论文摘要

近年来,随着人工智能(AI)、机器学习(ML)和深度学习算法的显著改进与发展,农业行业在精准性与准确性方面取得了重大进展,尤其是在棉花产业作物健康监测方面。然而,以往研究在泛化能力和计算成本方面存在一定局限。为应对这些挑战,本研究选取了十大深度学习框架——VGG16、VGG19、InceptionV3、Xception、ResNet50、MobileNetV2、ResNet101、ResNet152、ShuffleNetV2和EfficientNetB0,利用迁移学习对棉花叶片健康状况进行高精度分类。研究采用基于田间采集的SAR-CLD-2024数据集,该数据集包含涵盖七种不同叶片状况类别的9137张标准化图像。本研究系统考察了静态架构复杂度、动态训练收敛性与泛化性能之间的权衡关系。实验结果表明,即使在高度不平衡的真实测试环境中,结构设计与优化策略仍显著影响诊断效果。在所测试的模型中,EfficientNetB0取得了最高的平均准确率91.29%,ResNet50和ResNet101分别取得了90.48%和89.55%的较高平均准确率。相反,VGG19的总体准确率最低,为57.63%,表明在多类农业任务中架构简洁性与表征能力之间存在明显的权衡。此外,混淆矩阵和定性分析表明,早停策略对于缓解深层网络中的过拟合至关重要。除草剂药害和叶片斑驳等症状特征较易被捕获,而复杂的颜色与纹理叠加特征则与较浅层模型相比深层残差架构更高的误分类易感性相关。总体而言,本研究为开发低资源消耗的边缘导向智慧农业诊断系统提供了有价值的见解。

全文

Frontiers in Plant Science

Smart agriculture: deep
transfer learning frameworks
for high-accuracy cotton
crop disease classi fication
Quart-ul-ain Abro1,2, Junjie Xiang1*, Wazir Muhammad3*,
Mudasar Latif Memon4, Akash Kumar2,5 and Azhar Ali Laghari2
1School of Public Administration, Shanxi Agricultural University, Taigu, Shanxi, China, 2College of
Resources and Environment, Shanxi Agricultural University, Taigu, Shanxi, China, 3Department of
Biomedical Engineering, Balochistan University of Engineering and Technology, Khuzdar, Pakistan,
4Department of Computer Science, The University of Modern Sciences, Tando Muhammad Khan,
Sindh, Pakistan, 5Institute of Eco-Environment and Industrial Technology, Shanxi Agricultural
University, Taiyuan, China
In recent years, the agricultural industry has seen signi ficant progress in precision
and accuracy, due to the remarkable improvement and development of new
artificial intelligence (AI), machine learning (ML), and deep learning algorithms
especially for monitoring the crop health in cotton industry. However, previous
research has some limitations in terms of generalizability and computational cost.
These challenges have been addressed by considering the top ten deep learning
frameworks such as VGG16, VGG19, InceptionV3, Xception, ResNet50,
MobileNetV2, ResNet101, ResNet152, ShuffleNetV2 and EfficientNetB0 leveraging
transfer learning for high-accuracy classi fication of cotton leaf health conditions.
Utilizing the field-based SAR-CLD-2024 dataset comprising 9,137 standardized
images across seven distinct leaf condition categories. In this research, we
systematically examines the tradeoffs between static architectural complexity,
dynamic training convergence, and generalization performance. Experimental
results demonstrate that structural design and optimization strategies signi ficantly
affect diagnostic effectiveness, even within a highly imbalanced real-world test
environment. Among the models tested, Ef ficientNetB0 achieved the highest
mean accuracy of 91.29%, while ResNet50 and ResNet101 achieved strong mean
accuracies of 90.48% and 89.55%, respectively. Conversely, VGG19 yielded the
lowest overall accuracy of 57.63%, indicating a clear tradeoff between architec-
tural simplicity and representation capacity for multi-class agricultural tasks.
Furthermore, confusion matrix and qualitative analyses highlight that early stop-
ping was essential for mitigating over fitting in deeper networks. While features
such as herbicide damage and leaf variegation were readily captured, compli-
cated color and textural overlays correlated with higher misclassi fication vulner-
ability in shallower models compared to deeper residual architectures. Overall,
this work provides valuable insights for developing edge-oriented smart agricul-
ture diagnostic systems withlow resource consumption.
KEYWORDS
cotton leaf, plant disease detection, precision agriculture, smart agriculture, transfer
learning
Frontiers in Plant Science frontiersin.org01
OPEN ACCESS
EDITED BY
Yahui Guo,
Central China Normal University, China
REVIEWED BY
Mansi Sahu,
Uttaranchal University, India
Grace Shalini T,
SRM Institute of Science and Technology
(Deemed to be University) Research
Kattankulathur, India
*CORRESPONDENCE
Junjie Xiang
184096117@qq.com
Wazir Muhammad
wazirlaghari@buetk.edu.pk
RECEIVED 09 September 2026
REVISED 10 October 2026
ACCEPTED 19 October 2026
PUBLISHED 13 November 2026
CITATION
Abro Q-u-a, Xiang J, Muhammad W,
Memon ML, Kumar A and Laghari AA
(2026) Smart agriculture: deep transfer
learning frameworks for high-accuracy
cotton crop disease classi fication.
Front. Plant Sci. 17:1993878.
doi: 10.3389/fpls.2026.1993878
COPYRIGHT
© 2026 Abro, Xiang, Muhammad,
Memon, Kumar and Laghari. This is an
open-access article distributed under the
terms of the Creative Commons
Attribution License (CC BY).The use,
distribution or reproduction in other
forums is permitted, provided the
original author(s) and the copyright
owner(s) are credited and that the
original publication in this journal is
cited, in accordance with accepted
academic practice. No use, distribution
or reproduction is permitted which does
not comply with these terms.
TYPE Original Research
PUBLISHED 13 November 2026
DOI 10.3389/fpls.2026.1993878

1 Introduction
Cotton is a cash crop highly susceptible to bacterial, viral, fungal
and pest-related stresses that affect productivity and fiber quality,
making early detection for essential yield protection and sustainable
management (Anwar et al., 2023; Johri et al., 2024; Mudafale, 2026).
Manual inspection is slow, time-consuming, highly dependent on
the expert knowledge, and prone to misdiagnosis, especially when
similar symptom color, texture, or shape lead to delayed interven-
tion and sub optimal pesticide use ( Vallabhajosyula et al., 2022 ;
Jafar et al., 2024; Shafik et al., 2024). These constraints have boosted
the use of arti ficial intelligence (AI) in precision agriculture,
particularly deep learning for automated crop management and
disease tracking in field and edge settings ( Junaidi et al., 2025 ). For
instance, convolutional neural networks (CNNs) is the most pop-
ular approach among AI techniques for plant disease classi fication
due to its ability to learn discriminative representations directly
from images and minimize reliance on hand-crafted features
(Shrivastava et al., 2019 ; Attallah, 2023 ; Zhu et al., 2024 ). Transfer
learning is especially valuable in agriculture because labeled field
datasets are often limited, imbalanced, and heterogeneous, while
pretrained models can improve convergence and performance on
domain-specific tasks ( Ahmad et al., 2021 ). Previous research in
other crops has been found to be successful using pretrained models
like VGG16, VGG19, ResNet, Inception, Xception, DenseNet,
MobileNet and Ef ficientNet, although results are signi ficantly dif-
ferent depending on the dataset, symptom complexity, optimizer
used and model size ( Ngugi et al., 2024 ; Saleem et al., 2024 ). The
classification accuracy of transfer learning systems is high in cotton,
ranging from ensemble to single-model systems, with many eval-
uations performed on relatively small, curated, or geographically
limited datasets, which have limited comparability between systems
and the potential to cause problems when attempting to generalize
to real field conditions ( Nalwanga et al., 2025 ; Haque et al., 2026 ).
Moreover, cotton production is always at risk from many foliar
pathogens and environmental factors, which collectively make it
one of the most complicated phytosanitary issues in state-of-the-art
agriculture ( Chohan et al., 2020 ; Hossain et al., 2024 ). The most
harmful diseases include Bacterial Blight, Curl Virus, Herbicide
Growth Damage, Leaf Hopper Jassids, Leaf Reddening and Leaf
Variegation, all of which have speci fic and somewhat distinct
morphological features on the leaf surface that can be mistaken
for others, by even the most advanced agronomists ( Chaube and
Pundhir, 2005 ; Ghosal, 2022 ; Sain et al., 2024 ). Traditional diag-
nostic procedures for cotton diseases are mainly based on physical
field scouting of trained extension staff, followed by laboratory
culture or serological analysis, and expert diagnosis, which is costly
and time-consuming when applied on a large scale ( Arnal Barbedo,
2013; Dhavale et al., 2025 ). Furthermore, the subjective nature of
visual assessment can cause diagnostic confusion leading to delayed
action, misapplication of pesticides, and consequently worsening
losses ( Pethybridge and Nelson, 2015 ). The scarcity of plant
pathology resources is further intensi fied across numerous cotton
producing developing nations due to insuf ficient plant pathology
expertise, especially among resource-limited smallholder farmers
who possess minimal support systems when confronting dis-
ease challenges.
However, deep learning, and more precisely convolutional
neural networks (CNNs), has been shown to have transformative
potential to agricultural image analysis, providing automated,
objective and repeatable disease diagnosis based on leaf images at
high throughput ( LeCun et al., 2015 ; Kamilaris and Prenafeta-
Boldú, 2018). Motivated by the seminal success of CNNs in medical
image classification and natural scene recognition, the plant disease
detection community has increasingly embraced CNN-based pipe-
lines that can discriminate path ological leaf phenotypes with
accuracy surpassing human expert diagnosis. Transfer learning, in
particular, has emerged as the dominant paradigm for agricultural
imaging tasks, wherein pretrained feature representations acquired
from the large-scale ImageNet benchmark are fine-tuned on
domain-specific datasets, markedly reducing the labeled data re-
quirement and training time without compromising classi fica-
tion performance.
Though extensive research on CNN applications has been con-
ducted for plant diseases, there are few benchmarks for well-designed
CNN architectures based on the same balanced multi-class cotton
disease dataset that rigorously compare multiple state-of-the-art ar-
chitectures under standardized experimental conditions. Therefore,
previously proposed models can b e based on various experimental
contexts, which makes it difficult to compare and interpret thefindings
across studies, and can mislead practitioners looking for solutions that
will work in their specific deployment situations.
In this paper, a systematic benchmark transfer learning ap-
proach for automated multi-class cotton leaf disease classi fication
with SAR-CLD-2024 dataset is presented. The main contributions
of our proposed work in the field of precision agriculture and plant
pathology AI are as under:
 We evaluate ten state-of-the-art pretrained CNN architec-
tures on a seven-class cotton leaf health condition classi fi-
cation task, under a uniform experimental protocol.
 We report overall accuracy and precision, recall, and F1-
score of the top models for each disease class, showing
ResNet-50 and Ef ficientNetB0 to be the best with an overall
accuracy of around 91.29% and a balanced macro F1-score
of 0.91.
 We perform a comparative parameter and memory study of
all 10 architectures to show trade-off of computational cost
and classification performance to help select an architecture
to match requirements for resource-constrained deploy-
ment scenarios.
The remaining part of paper is arranged in the following way.
The literature related work on CNN-based plant disease detection is
surveyed in Section 2. Section 3 presents descriptions of the state-
of-the-art transfer learning models. The dataset, preprocessing
pipeline and methodology of the experiments are discussed in
Section 4. The experimental results and comparison are given in
Section 5. Finally, Section 6 concludes with a summary of the
findings and directions for future research.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org02

2 Related work
In recent years, the state of automation for detecting plant
diseases with the help of machine learning has made signi ficant
progress ( Saleem et al., 2021 ). Initial methods involved manually
engineered features including color histogram, local binary pattern,
and Scale-Invariant Feature Transform (SIFT) texture descriptors,
and used traditional classi fiers such as the support vector machines
(SVMs) and k-nearest neighbor (KNN) algorithms ( Pujari et al.,
2016). These approaches were shown to be feasible for controlled
laboratory datasets but performed badly when challenged by visual
variability in images collected in the field (Singh and Misra, 2017 ).
A major study by Upadhyay et al. (2025) paved the way by showing
that a deep learning model trained on ImageNet could achieve an
accuracy of 99.35% on the public dataset of 54,306 images, from 14
crop species and 26 diseases ( Ferentinos, 2018; Tugrul et al., 2022 )
validated that the VGG ’s and GoogLeNet variants always per-
formed better than traditional classi fiers in multi-class plant disease
recognition. Deep Learning is reshaping the cotton value chain,
automating processes such as crop monitoring, harvesting, and
grading cotton fiber. It has a high optimization potential in terms of
accuracy and ef ficiency in agriculture, but there are still certain
challenges, such as enormous computing demands and restricted
data generalizability. To realize sustainable automation in the
future, lightweight models, standardized datasets and integration
of multi-modal data are crucial ( Yang et al., 2025 ). Islam et al.
(2023) proposed a novel study known as a hybrid deep learning
system for the automatic detection of cotton leaf health conditions,
aiming to reduce the time and subjective aspects of traditional
methods through the application of the Ef ficientNetB3 and
InceptionResNetV2 architectures ( Kaur et al., 2025 ). Ahmed et al
(Ahmed, 2021 ), suggested to cope with the complicated image
processing and manual feature extraction that is necessary with
the existing method, this research proposes a custom Deep
Convolutional Neural Network (DCPLD-CNN) for automatic de-
tection of cotton diseases such as Bacterial Blight and rolling leaf
health conditions from images. Beside the custom model, the
research evaluates eight pre-tr ained architectures including
DenseNet121, ResNet50, and Xception fine-tuned via transfer
learning with extra dense layers. Patra et al ( Patra and Gajurel,
2025), report the major production threats from leaf-affecting
diseases in cotton crops, this study proposes an innovative, light-
weight deep learning framework that fills the gap for computation-
ally efficient agricultural tools. The architecture integrates a strategic
subset of trainable layers from the MobileNet model via transfer
learning, optimized using data augmentation, a learning rate decay
schedule, model checkpoints, a nd early stopping mechanisms.
Mulugu et al. (2025) , proposes a powerful deep learning system
based on DenseNet121 for automated classi fication of seven differ-
ent conditions of cotton leaves such as Bacterial Blight, Curl Virus,
Herbicide Growth Damage. Aslam et.al ( Aslam et al., 2025 ).,
addresses the challenges of multi-class cotton disease detection,
such as class imbalance, symptom variation, and the need for real-
time application by introducing a robust, automated deep learning
framework capable of classifying six distinct diseases. Chavan et al
(Chavan and Shirdhonkar, 2024 ), addresses the severe threat crop
diseases pose to agricultural production and the global food supply
by developing a real-time mobile application for automated cotton
disease detection, eliminating the need for specialized agricultural
expertise. Sajjan et al. (2025) , proposes a deep learning-based
framework for the classi fication of cotton plant diseases, which
uses a balanced custom dataset of cotton leaf images comprising
9000 images of nine different cotton plant diseases.
Nivethitha et al ( Nivethitha and Venkataramanan, 2026 ). pre-
sent a quantum-enhanced framework that combines wavelet pre-
processing, feature extraction with Ef ficientNet-B7 and Inception-
ResNet v2, the Archimedes Optimization Algorithm, and an
Attention-Based Quantum Convolutional Neural Network
(ABQCNN) tuned by the Adaptive Silver Fox Algorithm. The
framework achieves accuracies of up to 99.70% on tomato and
99.92% on cotton datasets, and Grad-CAM is used for interpreta-
bility. It may be noted that such a multi-stage pipeline increases the
design and tuning effort, which can limit its adoption in resource-
constrained agricultural settings. Kumar et al ( Kumar and
Thirupati, 2026 ). adapt DenseNet169 for seven cotton leaf condi-
tions on the SAR-CLD-2024 dataset of 9,137 images. A two-step
transfer learning strategy with fine-tuning of 161 layers and
enhanced data augmentation results in a validation accuracy of
96.83% and a macro-average AUC of 99.81%. However, their study
considers a single backbone, so a fair comparison across different
architectures under identical training settings is not available.
Transfer learning has also been applied to other crops with
limited image data. Mazumder et al. (2024) propose LeafDoc-Net,
which combines DenseNet121 and MobileNetV2 with attention
mechanisms, extra dense layers, swish activation, and batch nor-
malization. LeafDoc-Net outperforms existing methods on cassava
and wheat datasets, while Grad-CAM++ explains its decisions.
Verma et al. (2026) introduce SLDCNet, a customized
EfficientNetB3 model for foliar diseases of Solanaceous vegetables.
SLDCNet obtains a testing accuracy of 99.03% on 5,211 field images
of nine disease classes. Shrikhande et al ( Shrikhande and Gawade,
2026). compare classical machine learning, custom CNNs, and
transfer learning on the PlantVillage dataset of 38 classes. The
accuracy of SVC-RBF peaks at 83.58% and that of the best custom
CNN reaches 95.71%, whereas MobileNetV2 achieves 99.30%. The
authors further extend this baseline to a multimodal decision-
support framework that fuses image, soil, and weather data.
On the other hand, a number of recent works combine
convolutional backbones with Vision Transformers (ViT) to cap-
ture both local lesion texture and long-range context. Potharaju
et al. (2026) apply ViT-B/16 with transfer learning to bean leaf
diseases and obtain a validation accuracy of 97.52% on the I-Bean
dataset, with Grad-CAM++ heatmaps highlighting the diseased
regions. Ali et al. (2026) integrate Ef ficientNet and ViT for the
augmented PlantVillage dataset of 61,486 images. The hybrid model
achieves 93% accuracy, as com pared to 89% for standalone
EfficientNet and ResNet, and augmentation adds a further 4%
gain. Similarly, Jawed et al. (2026) combine Ef ficientNet-B7 and
ViT-B16 and report an accuracy of 98.13% on 21,534 images of 38
classes. Chellamani et al. (2026) follow a two-phase design in which
pretrained CNNs serve as feature extractors and a ViT is paired with
classical classi fiers. The MobileNetV2-SVM combination reaches
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org03

89.44%, while the ViT-SVM framework reaches 98%. The reason
for the large spread among these figures lies in the fact that each
study uses a different dataset, split, and augmentation protocol.
Hence, the reported accuracies cannot be compared directly.
Another line of work addresses deployment under field condi-
tions. Rajendran et al. (2026) propose an Edge-AI crop protection
system that integrates sensor fusion, UAV hyperspectral sensing,
ViTs, Graph Neural Networks, and Federated Learning. The system
detects diseases with an accuracy of 98.99% and forecasts pests with
an accuracy of 96.34%, while reducing pesticide use by 32%. Wang
et al. (2026) build a solar-powered edge prototype for rice pest
detection and trapping. It attains 96.1% mAP@0.5 at 59.4 FPS on a
Jetson Xavier NX with TensorRT. Haqiq et al. (2026) present
StrawberryDualNet, a lightweight CNN –Transformer model of
only 0.04M parameters (0.72 MB), which achieves 95.1% accuracy
and runs at 83 FPS under 16-bit quantization. In a scoping review
of 224 studies, Abayeva et al ( Abayeva and Kabdrakhmanova,
2026). observe that most evidence rests on single-season validation,
which overstates the portability of digital technologies to real fields.
To summarize, existing studies report very high accuracies on
cotton and other crops, but they differ in datasets, backbones, and
training protocols. Moreover, only a few of them examine
several architectures on one cotton dataset under the same
settings. Motivated by these observations, this work evaluates ten
models on the SAR-CLD-2024 dataset with a uniform train-
ing con figuration.
3 Transfer learning models
Transfer learning refers to the practice of initializing a target-
domain model with parameters learned during training on a large-
scale source domain, subsequently adapting those parameters to the
target task through fine-tuning on domain-speci fic labeled data
(Cui et al., 2018 ; Gholizade et al., 2025 ). In the context of plant
disease image classification, source models are invariably pretrained
on ImageNet, enabling pretrained convolutional layers to encode
rich hierarchical feature representations that transfer effectively to
fine-grained botanical image discrimination ( Dong et al., 2023; Guo
et al., 2025 ). The following are succinct descriptions of each
architecture evaluated in this study.
3.1 EfficientNetB0
In Ef ficientNetB0, a principled multipart scaling approach is
used to mutually scale the network depth, width and input size
resolution ( Tan and Le, 2019 ) as illustrated in Figure 1. It is built
with mobile inverted bottleneck convolution (MBConv) blocks with
squeeze-and-excitation attention, which leads to a high accuracy-
to-parameter ratio, suitable for transfer learning tasks on lim-
ited datasets.
3.2 InceptionV3
InceptionV3 uses parallel convolutional branches of different
filter sizes of an order 1×1, 3×3, and 5×5 within the same inception
module which allow multi-scale features extraction ( Szegedy et al.,
2016). Factorized convolutions and batch normalization lower the
computational cost without compromising representational capac-
ity as shown in Figure 2.
3.3 MobileNetV2
MobileNetV2 implements depth-wise separable convolutions
and inverted residual structures with linear bottlenecks for place-
ment on mobile and embedded devices ( Sandler et al., 2018 )a s
shown in Figure 3. It has a high representation capability within a
low parameter budget.
3.4 ResNet-50, ResNet-101, and ResNet-
152
The vanishing gradient problem is solved by Residual Networks
(He et al., 2016 ) using identity shortcut connections. The difference
between ResNet-50, ResNet-101, and ResNet-152 lies on the depth
and the use of bottleneck residual blocks, which are triplets of 1×1,
3×3, and 1×1 convolution as shown in Figure 4. Deeper variants are
associated with learning more abstract semantic representations
that align with fine-grained discrimination of diseases.
3.5 ShuffleNetV2
ShuffleNetV2 uses channel splitting and channel shuf fling
operations to facilitate the ef ficient flow of information while
FIGURE 1
General architecture of Effi cientNetB0 for cotton crop disease classi fication.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org04

reducing the cost of operations in memory ( Ma et al., 2018 )a s
shown in Figure 5 . It is the most computation parsimonious
architecture in this benchmark, and relevant to edge deploy-
ment scenarios.
3.6 VGG16 and VGG19
The VGG designs use completely homogeneous ones with only
3×3 convolutional filters, with a depth of 16 layers of weight in the
first one, and 19 layers in the second one (Simonyan and Zisserman,
2014) as shown in Figure 6. Although they perform well on large-
scale benchmarks, they can be inef ficient in fine-tuning on small
multi-class datasets due to their large fully-connected parame-
ter counts.
3.7 Xception
Xception is an improvement of Inception by substituting the
Inception modules with depth-wise separable convolutions, which
split the channel-wise and spatial convolutions into independent
ones (Chollet, 2017) as shown in Figure 7. This factorization is used
to decrease parameters and enh ance cross channel correla-
tion learning.
4 Materials and methods
4.1 Dataset description
The experiments in this study utilize the publicly available SAR-
CLD-2024 benchmark dataset ( Bishshash et al., 2024 ), designed for
cotton leaf condition classi fication. The dataset comprises high-
resolution images captured under diverse field conditions (including
varying illumination, angles, and leaf development stages) categorized
across seven distinct fine-grained classes: Bacterial Blight, Curl Virus,
Healthy Leaf, Herbicide Growth Damage, Leaf Hopper Jassids, Leaf
Reddening, and Leaf Variegation. The original unaugmented bench-
mark dataset consists of 2,137 images, with the following exact class-
wise distribution illustrated in the testing distribution: Leaf
Reddening contains 578 samples, Curl Virus contains 431 samples,
Herbicide Growth Damage contains 280 samples, Healthy Leaf
contains 257 samples, Bacterial Blight contains 250 samples, Leaf
Hopper Jassids contains 225 samples, and Leaf Variegation contains
116 samples. As evidenced by these counts and the right-hand panel
of Figure 8 , the raw dataset exhibits a notable class imbalance for
instance, Leaf Reddening contains nearly five times more samples
than Leaf Variegation. Training deep learning architectures directly
on such imbalanced data can bias the model toward majority classes.
FIGURE 2
General architecture of InceptionV3 for cotton crop disease classi fication.
FIGURE 3
General architecture of MobileNetV2 for cotton crop disease classi fication.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org05

FIGURE 4
General architecture of ResNet-50, ResNet-101, and ResNet-152 for cotton crop disease classi fication.
FIGURE 5
General architecture of shuf fleNetV2 for cotton crop disease classi fication.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org06

To overcome this limitation and ensure robust feature representation,
the data pipeline was structured through a strict separation and
augmentation protocol. First, the original unaugmented dataset was
partitioned into training and testing sets prior to any synthesis
operations to ensure complete independence. Second, the testing
partition contains solely original, untouched, and completely unseen
images used exclusively as a validation set. No augmented counter-
part, derivative, or transformed version of any test image was ever
introduced into the training phase, ensuring that the evaluation
framework remains robust and entirely free from data leakage.
Third, data augmentation techniques including rotation, shifting,
shearing, and horizontal/vertical flipping were applied exclusively to
the training subset until every category reached an equalized thresh-
old. As visually detailed in the left-hand panel of Figure 8 (“Training
(Augmented) Class Distribution ”), this synthesis process yielded a
perfectly balanced training corpus totaling 7,000 augmented images
with a uniform flat distribution of exactly 1,000 balanced training
instances per class across all seven categories, ensuring equal gradient
contributions during backpropagation. Across the balanced training
corpus and the untouched testing partition, the framework encom-
passes a total of 9,137 high-resolution images. The final class
distribution for both the training and testing sets is comprehensively
contrasted and illustrated in Figure 8.
4.2 Data preprocessing and augmentation
For all images were resized to the input size of the respective
architecture such as 224×224 pixels for Ef ficientNetB0,
MobileNetV2, ResNet variants, Shuf fleNetV2 and VGG models,
299×299 pixels for InceptionV3 and Xception. ImageNet channel-
wise-mean subtraction and standard deviation scaling was per-
formed on the pixel values. The augmentation during training was
done through the Keras ImageDataGenerator pipeline and in-
cluded horizontal flipping of the image, rotation, width shift
range of 0.2, height shift range of 0.2, zoom range of 0.2, and
shear range of 0.2.
4.3 Model architecture and fine-tuning
protocol
A transfer learning benchmarking framework was developed for
the classifi cation of cotton leaf health condition based on multi-
backbone models, as illustrated in Figure 9. ImageNet weights
served as the initial weights for each backbone, and all transfer
learning models were utilized as feature extractors with their
internal layers frozen during the initial extraction phase, leaving
only the custom classi fication head trainable. The custom classifi -
cation head consolidates global average pooling, a dropout regu-
larization layer (rate of 0.2), an intermediate dense layer of 128 units
using the Rectified Linear Unit (ReLU) activation function, a second
dropout layer (rate of 0.5), and a final classifi cation output dense
layer of 7 units utilizing the softmax activation function. Training
was conducted using Python 3.12 and TensorFlow version 2.20.0
with the integrated Keras API in a Google Colab cloud environment
using a GPU-accelerated runtime. To ensure deterministic behavior
and exact reproducibility, random seeds for both TensorFlow and
NumPy were explicitly set to 42 for the primary run and 42-46 for
the multi- seed evaluation. All models were trained for a maximum
of 30 epochs with a batch size of 32 using the Adam optimizer with
an initial learning rate of 0.0001, optimized with categorical cross-
entropy loss. Training was further supported by callbacks including
early stopping (monitoring validation accuracy with a patience of 5
epochs and restore_best_weights=True), ModelCheckpoint, and
ReduceLROnPlateau (factor=0.2, patience=5).
FIGURE 6
General architecture of VGG16 and VGG19 for cotton crop disease classi fication.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org07

4.4 Performance evaluation metrics
Model performance was evaluated on the 2,137-image test set
using precision, recall, F1-score, and multi-class accuracy. Multi-
class accuracy is calculated as the total number of correctly
predicted samples across all seven classes divided by the total
number of evaluation samples as Equation 1 :
Accuracy = o
C
i=1TPi
N (1)
Precision for each class i is the proportion of samples predicted
as class i that actually belong to it ( Equation 2).
Precisioni = TPiTPi + FPi (2)
To summarize this across classes, the macro-averaged preci-
sion is the unweighted mean of the per-class precision values
(Equation 3 ).
Macro − Precision = 1C oCi=1 TPiTPi + FPi (3)
Recall for each class is the proportion of true class-i samples that
the model correctly identifies (Equation 4), and the macro-averaged
recall is the unweighted mean of these values across all classes
(Equation 5).
Per − Class Recall Recalli = TPiTPi + FNi (4)
Macro − Averaged Recall = 1C oCi=1 TPiTPi + FNi (5)
The per-class F1-score is the harmonic mean of precision and
recall (Equation 6), and the macro-averaged F1-score is obtained by
averaging the per-class F1-scores ( Equation 7).
Per − Class F1−Score(F1i)=2 /C2 Precisioni /C2 RecalliPrecisioni + Recalli
= 2 TPi
2 TPi + FPi + FNi
(6)
Macro–Averaged F1–Score(Macro /C0 F1) = 1C oCi=1F1i (7)
where, TPi (True Positives) is the number of samples correctly
predicted as class i, FPi (False Positives) is the number of samples
incorrectly predicted as classi (they actually belong to another class),
TPi + FPi represents the total number of samples predicted as class i
by the model (the column sum for class i in a confusion matrix) and
C represents the total number of cotton leaf disease classes.
5 Experimental results
5.1 Training and testing dataset
The proposed dataset ( Bishshash et al., 2024 ) is a massive field
acquired dataset that includes cotton leaf images of both healthy
FIGURE 7
General architecture of Xception for cotton crop disease classi fication.
FIGURE 8
Class distribution of the SAR-CLD-2024 dataset training and testing partition counts across the seven cotton leaf disease categories.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org08

leaves and various types of diseases. By using this dataset for the
automatic identifi cation and classifi cation of cotton leaf diseases,
researchers were able to achieve a high accuracy of about 96% with
high precision scores, a major step forward in the agriculture sector.
5.2 Training performance and convergence
analysis
To highlight the learning curves of ten different deep learning
architectures such as VGG16, VGG19, InceptionV3, Xception,
ResNet50, MobileNetV2, ResNet101, ResNet152, Shuf fleNet, and
EfficientNetB0 with regard to training/validation accuracy and
loss curves.
The learning rate was set to 0.0001 for training, with the
maximum training epochs set at 30, but an early stopping technique
was effectively used as a regularization technique to avoid
overfitting. This mechanism stopped the training when the loss
during validation stopped decreasing further, with the training
sessions taking anywhere from about 11 epochs to 24 epochs
depending on the model.
The training and validation curves response analysis indicates
different categories of behavior in the architectures, with one group
exhibiting very stable converging behavior and good generalization.
VGG16, VGG19, and MobileNet are all part of this benign set of
models, with a training versus validation loss difference that is very
small. In the case of VGG16 and VGG19, their training loss
becomes very close to the validation loss till the minimum, and
VGG16 stops at epoch 24 while VGG19 has a rapid convergence
and reaches its minimum at epoch 17.
MobileNetV2 also achieves a Plateau of accuracy around 85%
quite early, has validation curves that closely follow training curves
and is one of the best balanced and most consistent models and
there is no signifi cant variance. The second group, which includes
InceptionV3, Xception, and the deep ResNet family (ResNet50,
ResNet101 and ResNet152), shows a clear generalization gap and
overfitting tendencies. The larger the depth of the ResNet architec-
tures, the more pronounced this difference becomes, as an accuracy
of 90% is achieved for training.
Importantly, the validation loss curves of these models start to
level off around epochs 4-6, and start a subtle upward or flat trend as
the training loss continues to decrease. Xception and InceptionV3
are similar, achieving training accuracy greater than 85%, but the
validation accuracy only reaches 65% to 70% and early stopping
callback is activated between epoch 14 and 16, stopping them from
getting a lot worse.
Last but not least, the lightweight architectures ShuffleNetV2 and
EfficientNetB0 offer a distinctive optimization, and volatility, during
training. Finally, ShuffleNetV2 exhibits the largest training instability
among the cohort, with an extremely jagged and noisy validation loss
curve, that oscillates sharply before early stopping ends the run early
around epoch 23, indicating a highly non-convex optimization
landscape or a slightly too high learning rate. However,
EfficientNetB0 hit the stopping criteria very early at epoch 11,
although it achieved high baseline accuracy quickly, its validation
loss did appear to bounce and pop back up, causing an early halt. In
all ten models, early stopping with restoration of the bestweights was
a critical training safeguard that preventedoverfitting during contin-
ued training up to the 30-epoch limit as shown in Figure 10.
5.3 Quantitative comparative analysis of
model in terms of confusion matrices
For the model evaluated on the 7-class plant pathology task, the
quantitative analysis is done using the ten different deep learning
architectures included as VGG16, VGG19, InceptionV3, Xception,
ResNet50, MobileNetV2, ResNet101, ResNet152, Shuf fleNet and
EfficientNetB0. The target categories are as a Bacterial Blight, Curl
Virus, Healthy Leaf, Herbicide Growth Damage, Leaf Hopper
Jassids, Leaf Reddening, and Leaf Variegation.
In this analysis, the actual distribution of the labels in each of
these matrices are compared with the distribution of the labels
predicted by the model, thereby providing important insights into
the performance of the structural model, into shared morphology
among speci fic diseases, and into the overall robustness of the
classification across various granularities of the network.
In both Herbicide Growth Damage and Leaf Variegation, the
classes are easily interpreted in all ten models. The true positive
values taken along the main diagonal of confusion matrices are very
high and consistent, ranging from 253 to 275, with very little
leakage into other categories for Herbicide Growth Damage. This
FIGURE 9
Smart agriculture-based deep learning frameworks for high-accuracy cotton crop disease classi fication.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org09

means that herbicide damage causes very distinct structural or color
abnormalities, e.g. uniform chlorosis or geometric leaf distortion,
which are readily distinguishable from biological infections by the
deep networks.
Likewise, Leaf Variegation performs very well on all the data,
consistently with 101 to 116 true positive and low number of false
positive/false negative, largely due to the highly contrasted, pat-
terned and striking nature of v ariegation, which is an easily
recognizable visual signature. The most interesting diagnostic
issue in this dataset is identifying Leaf Reddening; the evolution
of the models across the time series of comparisons reveals a
considerable evolutionary jump in the way the various architectures
deal with this particular feature.
In simpler sequential models, VGG16 and VGG19, there are
118 misclassifi ed Leaf Reddening instances as Bacterial Blight and
VGG19 misclassi fies 144 instances of Leaf Reddening as Bacterial
Blight and another 119 as Curl Virus. This large off-diagonal
spreading implies that shallower or non-residual architectures are
unable to separate Bacterial Blight and Curl Virus from thegeneral
coloration of reddening leaves, which is more indicative of other
reddening processes of reddening such as shallower and non-
residual architectures. This confusion is increasingly reduced
toward inception-based networks and deeper residual networks.
The misclassi fication of Leaf Reddening to Bacterial Blight is
significantly reduced to 89 and 39 cases by InceptionV3 and
Xception, respectively. The ResNet family (ResNet50, ResNet101,
and ResNet152) offers the best performance for Leaf Reddening,
with up to 519 correct classi fications in ResNet50, 509 in
ResNet101, and 508 in ResNet152. They drastically suppress off-
diagonal leakage, demonstrating that architectures featuring skip
connections and deeper topologies achieve higher fidelity in sepa-
rating classes without being overly reliant on global color cues.
Moreover, the classi fication rate of Curl Virus increases with the
increasing complexity and depth of the model, ranging from 236 in
VGG16 to 368 in ResNet101 and 365 in ResNet152. The lightweight
architectures, ShuffleNetV2 and EfficientNetB0 can achieve reason-
able true positive diagonals with a small compromise on fine-
grained classi fication boundary resolution, as seen in the slight
bleeding between multiple classes, especially for Healthy Leaf and
Leaf Hopper Jassids as shown in Figure 11 .
5.4 Qualitative performance and visual
inference analysis
All ten deep learning models were evaluated in the qualitative
performance analysis and localized class inferences on the cotton
crop leaf pathology images as shown in Figure 12 . The samples in
each sub-figure display a grid of test leaves with the predicted class
label and the prediction con fidence score in percentage. These
visual examples give a clear localized view of how each architecture
deals with leaf geometry, texturing, macro-coloration and position-
ing of necrotic spots in the formation of the final output. One
striking pattern in these figures is that the greater one ’s confidence
in the prediction, the more complex the network architecture.
The con fidence performed very well with deeper residual
networks like ResNet50, ResNet101 and ResNet152, sometimes
exceeding 95% for very distinguishable classes like Herbicide
Growth Damage and Leaf Reddening. This suggests that the deep
residual models have greater con fidence in the predictions, which is
linked to having less uncertainty within the deep network, and
aligning more closely with the target diagnosis. The score distribu-
tions and the baseline con fidences of the shallower sequential
models VGG16 and VGG19, on the other hand, are far broader,
and some correct classi fications have con fidence below 65%, such
asLeaf Hopper Jassids (62.72%) and Leaf Variegation (39.51%).
The qualitative results across these images show the reasons for
the separation of certain classes being as clean as it is. The leaves are
highly dramatic, elongated, sharp in structural distortion and severe
FIGURE 10
Training loss and accuracy performance curves of different transfer
learning models for cotton crop disease classi fication.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org10

geometric narrowing, which is why there is a high to complete
confidence with herbicide Growth Damage across different models
such as ResNet50, ResNet101, and Ef ficientNetB0. This morpho-
logical deformation is not caused by biological pathogens, which
instead cause a speckled or rounded damage, and is therefore very
difficult to be confused by a biological cause, for which reason it is
recorded by almost all models with the highest certainty.
Likewise, the leaf variegation is constructed into high contrast
patchwork of white, light green and dark green sections within an
individual leaf. The confidence values in models along the spectrum
are able to capture these global color boundaries with ease, though
there is a wide variation (54.99 –99.49%) in the lightest architec-
tures, such as Shuf fleNetV2 and Ef ficientNetB0, which may be
attributable to the lower spatial resolution of compressed channels.
FIGURE 11
Quantitative comparative analysis of model in terms of confusion matrices.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org11

Moreover, the complexity of the diagnosis is shown in
Figure 12 , such as the Curl Virus and Leaf Reddening. Curl Virus
symptoms include severe 3D warping with curling and folding of
the leaf blade while Leaf Reddening symptoms are deep purple or
crimson coloration spreading from the veins. The VGG models
exhibit lower con fidence in the presence of leaves with local viral
curling but a higher level of certainty for deeper architectures when
dealing with these complex spatial transformations. The prediction
grids are an important qualitative validation tool for automated
crop monitoring. In summary, ResNet, Ef ficientNetB0 and
MobileNetV2 give the most reliable and clear-cut diagnostic results
and hence they are considered as the best candidates for real-world
computer vision based field applications in agriculture.
5.5 Misclassi fication pattern analysis
The top ten models ’ most prevalent misclassi fication patterns
can be observed in Figure 13 . Across all the AI models, the most
common mistake is confusing Leaf Reddening with Leaf Hopper
Jassids because the two conditions look very similar in real life, as
the yellowing caused by Jassid bugs closely resembles the reddening
caused by Leaf Reddening under natural sunlight. In addition, many
models often misidentify Curl Virus as Leaf Hopper Jassids, which
is the second most common misidenti fication. This demonstrates
the difficulty in distinguishing between these cotton conditions and
thus highlights the need for further research to explore the use of
advanced attention techniques to examine the plant veins and leaf
pores to minimize these visual overlaps in the future.
5.6 Architectural complexity and dynamic
convergence analysis
The Figure 14 provide an extensive dual-perspective analysis of
ten different deep learning models, and link the static architectural
footprint in terms of total number of parameters and memory
allocation size of their dynamic training loss and accuracy conver-
gence with early stopping epochs. The architectural complexity
chart arranges the models in an orderly ascending fashion of
complexity, beginning with the most lightweight models
(ShuffleNetV2 with 1.4 million parameters and 5.3 MB memory
requirement), moving through semi-heavy models (such as VGG16
with 14.8 million parameters and 56.4 MB memory requirement),
and ending with the heaviest models (ResNet152 with 58.6 million
parameters and a massive 223.7 MB memory requirement).
The architectural hierarchy, coupled with the associated learn-
ing curve, sheds light on the direct operating compromises between
model size, training stability, generalization and the importance of
early stopping in optimizing resource-heavy training pipelines. One
major result of this comparison is that the relationship between the
size of the architecture and the ef ficiency of learning is not linear, so
that the amount of parameters in the architecture is not always the
most important factor when setting up the boundaries for learning
generalizations. Indeed, the extremely lightweight MobileNetV2
(3.4M parameters, 12.8 MB memory) and Ef ficientNetB0 (4.2M
parameters, 16.1 MB memory) converge very quickly and have
comparable accuracy plateaus and Ef ficientNetB0 stops early at
epoch 11 because it is optimized very quickly.
On the other hand, the much larger sequential networks,
VGG16 and VGG19, have greater number of parameters and
have smoother learning gradients but slower. The results for
VGG16 are also obtained after 24 epochs, with very close difference
between training and validation results, as the simpler, uniform
depth model does not suffer the dramatic generalization error that is
observed in more complex parameter models. Important, the
deeper models in the class show a clear structural over fitting
susceptibility that necessitates a strong regulatory measure. The
ResNet family (ResNet50, ResNet101, ResNet152) and InceptionV3
and Xception show another pronounced increase in the generali-
zation gap as the number of parameters gets larger than 20 million.
In these models, the training accuracy continuously gets closer to
90% and training loss continuously gets closer to 0, but the
validation loss curve either levels off or even peaks early during
training, typically at epochs 4-6. Finally, by using architectural
metric mapping and convergence plots it is evident that larger
FIGURE 12
Qualitative comparison analysis of top ten models.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org12

models are capable of large capacity but have signi ficant validation
divergence which must be addressed via early stopping. At the same
time, optimized compact networks, such as MobileNetV2, allow for
highly balanced and resource ef ficient convergence to be achieved,
making them highly attractive candidates for deployment in re-
source and hardware limited clinical or industrial settings.
5.7 Quantitative evaluation of deep
learning models for plant disease
classification
The performance of ten different deep learning architectures,
including a VGG16, VGG19, InceptionV3, Xception, ResNet50,
FIGURE 13
Most common inter-class misclassi fication patterns for the ten architectures on the SAR-CLD-2024 test set.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org13

MobileNetV2, ResNet-101, ResNet-152, Shuf fleNetV2 and
EfficientNetB0, was tested. The model were training on 7,000
class-balanced augmented images and tested on the 2,137 original,
untouched test images covering seven leaf conditions (Bacterial
Blight, Curl Virus, Health Leaf, Herbicide Growth Damage, Leaf
Hopper Jassids, Leaf Reddening, Leaf Variegation).
The overall classi fication accuracy analysis showed that
EfficientNetB0 is the best model, with peak accuracy of 0.91,
macro F1 score of 0.91 and weighted F1 score of 0.91.
EfficientNetB0 showed outstanding performance in all the catego-
ries, including the perfect recall of 1.00 for Leaf Variegation, and a
good performance on the classes that had historically
been challenging.
The ResNet family were also strong performers with ResNet50
and ResNet101 achieving an overall accuracy of 0.90 and macro and
weighted F1 of 0.90. To further improve performance, the authors
tried to scale up to a deeper architecture, but they did not see any
further improvement; ResNet152 saw a marginal drop in overall
accuracy to 0.89 and a macro F1-scoring of 0.88, indicating that
further depth did not yield p roportional gains for this
dataset’s complexity.
Xception achieved an overall mean accuracy of 83.09% and a
weighted F1-score of 0.84, closely followed by InceptionV3 with an
accuracy of 78.88% and an F1-score of 0.80. The lightweight
MobileNetV2 architecture achieved a mean accuracy of 87.59%
with balanced F1-scores, while Shuf fleNetV2 recorded a moderate
accuracy of 73.30%. Conversely, the lowest performance was
consistently observed in legacy VGG architectures. VGG16 ob-
tained a mean accuracy of 61.88%, while VGG19 recorded the
lowest performance with a mean accuracy of 57.63% and a weighted
F1-score of 0.59. Performance variations across classes highlight
significant differences in discriminative difficulty. Herbicide Growth
Damage was the most readily identifi able condition across nearly
the entire architectural spectrum, with InceptionV3, MobileNetV2,
ResNet50, ResNet101, ResNet152, and Ef ficientNetB0 achieving a
precision of 1.00 with zero false positives.
On the other hand, Healthy Leaf and Leaf Hopper Jassids were
the most dif ficult group to categorize. For example, the F1 score of
VGG19 was 0.38 for a Healthy Leaf and 0.43 for Leaf hopper Jassids,
indicating that they were very poor at detecting the subtle, fine-
grained visual features that characterize them. These classifi cation
issues were resolved to a signifi cant degree by the advanced models,
however; Ef ficientNetB0 signi ficantly improved the F1-scores for
Healthy Leaf and Leaf Hopper Jassids, at 0.92 and 0.81, respectively
in Table 1. Overall, the findings underscore the promise of using
EfficientNetB0 and ResNet models for achieving high accuracy in
automated diagnostic tasks in this field.
5.8 Performance analysis of deep learning
models for plant disease detection
The results of the seven different plant health categories reveal
that the performance of the different deep learning architectures are
widely different across the seven categories. As can be seen in
Table 2, the error rates are the highest for the early sequential
networks, especially VGG16, reaching 49.13% error rate for Leaf
Reddening, and VGG19 reaching a signifi cant 59.92% error rate for
Healthy Leaf. These models show that there is a clear lack of ability
to recognize fine-grained and complex features but there is a relative
success in recognizing visual signatures such as Herbicide Growth
Damage and Leaf Variegation. Multi-scale feature extraction net-
works, on the other hand, provide a signi ficant performance
improvement such as an InceptionV3 is able to achieve a 14.38%
error rate for Curl Virus, and Xception applies depthwise separable
convolutions to reach an error rate of 9.33% for Leaf Hopper
Jassids. The incorporation of residual skip connections coincides
with lower feature degradation, yielding state-of-the-art error rates,
such as only 2.33% for ResNet50 on Healthy Leaf and 1.79% on
Herbicide Growth Damage, as shown in the same Table 2.
Deeper variations such as ResNet101 and ResNet152 exhibit
very robust and very low error rates in all the cases, especially when
it comes to more complex path ological patterns like Leaf
Reddening, whose error rates are around 11% to 12%. When
comparing lightweight networks, MobileNetV2 demonstrates
highly competitive edge performance with error rates of 2.14%
and 3.11% on Herbicide Growth Damage and Healthy Leaf,
FIGURE 14
Computational complexity of the ten models in terms of number of parameters (millions) and memory size (MB).
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org14

TABLE 1 Class-wise and overall performance metrics (precision, recall, F1-score) of deep learning models on the leaf condition dataset.
Model Classi fication report Precision Recall f1-score Support
VGG16 Bacterial Blight 0.51 0.66 0.58 250
Curl Virus 0.74 0.55 0.63 431
Healthy Leaf 0.41 0.56 0.47 257
Herbicide Growth Damage 0.96 0.97 0.97 280
Leaf Hopper Jassids 0.42 0.61 0.49 225
Leaf Reddening 0.77 0.51 0.61 578
Leaf Variegation 0.73 0.95 0.83 116
accuracy 0.64 2137
macro avg 0.65 0.69 0.65 2137
weighted avg 0.68 0.64 0.64 2137
VGG19 Bacterial Blight 0.42 0.71 0.53 250
Curl Virus 0.59 0.58 0.59 431
Healthy Leaf 0.37 0.40 0.38 257
Herbicide Growth Damage 0.94 0.95 0.94 280
Leaf Hopper Jassids 0.41 0.44 0.43 225
Leaf Reddening 0.79 0.43 0.56 578
Leaf Variegation 0.66 0.94 0.77 116
accuracy 0.59 2137
macro avg 0.60 0.64 0.60 2137
weighted avg 0.63 0.59 0.59 2137
InceptionV3 Bacterial Blight 0.68 0.75 0.72 250
Curl Virus 0.84 0.86 0.85 431
Healthy Leaf 0.80 0.80 0.80 257
Herbicide Growth Damage 1.00 0.90 0.95 280
Leaf Hopper Jassids 0.53 0.88 0.66 225
Leaf Reddening 0.94 0.65 0.77 578
Leaf Variegation 0.79 0.94 0.86 116
accuracy 0.80 2137
macro avg 0.80 0.83 0.80 2137
weighted avg 0.83 0.80 0.80 2137
Xception Bacterial Blight 0.78 0.72 0.75 250
Curl Virus 0.90 0.86 0.88 431
Healthy Leaf 0.83 0.86 0.85 257
Herbicide Growth Damage 0.99 0.96 0.97 280
Leaf Hopper Jassids 0.55 0.91 0.68 225
Leaf Reddening 0.92 0.74 0.82 578
Leaf Variegation 0.83 0.87 0.85 116
accuracy 0.83 2137
macro avg 0.83 0.85 0.83 2137
weighted avg 0.86 0.83 0.84 2137
ResNet50 Bacterial Blight 0.87 0.79 0.83 250
Curl Virus 0.98 0.84 0.91 431
Healthy Leaf 0.79 0.98 0.88 257
Herbicide Growth Damage 1.00 0.98 0.99 280
Leaf Hopper Jassids 0.74 0.93 0.83 225
(Continued)
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org15

TABLE 1 Continued
Model Classi fication report Precision Recall f1-score Support
Leaf Reddening 0.95 0.90 0.92 578
Leaf Variegation 0.95 0.92 0.93 116
accuracy 0.90 2137
macro avg 0.90 0.91 0.90 2137
weighted avg 0.91 0.90 0.90 2137
MobileNetV2 Bacterial Blight 0.70 0.76 0.73 250
Curl Virus 0.95 0.91 0.93 431
Healthy Leaf 0.80 0.97 0.88 257
Herbicide Growth Damage 1.00 0.98 0.99 280
Leaf Hopper Jassids 0.70 0.91 0.79 225
Leaf Reddening 0.92 0.73 0.81 578
Leaf Variegation 0.91 0.92 0.91 116
accuracy 0.86 2137
macro avg 0.85 0.88 0.86 2137
weighted avg 0.87 0.86 0.86 2137
ResNet101 Bacterial Blight 0.81 0.83 0.82 250
Curl Virus 0.96 0.85 0.90 431
Healthy Leaf 0.85 0.98 0.91 257
Herbicide Growth Damage 0.99 0.96 0.98 280
Leaf Hopper Jassids 0.79 0.89 0.84 225
Leaf Reddening 0.92 0.88 0.90 578
Leaf Variegation 0.90 0.96 0.93 116
accuracy 0.90 2137
macro avg 0.89 0.91 0.90 2137
weighted avg 0.90 0.90 0.90 2137
ResNet152 Bacterial Blight 0.85 0.76 0.80 250
Curl Virus 0.98 0.85 0.91 431
Healthy Leaf 0.84 0.96 0.90 257
Herbicide Growth Damage 1.00 0.97 0.98 280
Leaf Hopper Jassids 0.67 0.93 0.78 225
Leaf Reddening 0.93 0.88 0.90 578
Leaf Variegation 0.92 0.92 0.92 116
accuracy 0.89 2137
macro avg 0.88 0.90 0.88 2137
weighted avg 0.90 0.89 0.89 2137
ShuffleNetV2 Bacterial Blight 0.62 0.60 0.61 250
Curl Virus 0.96 0.59 0.73 431
Healthy Leaf 0.59 0.76 0.66 257
Herbicide Growth Damage 0.84 0.96 0.90 280
Leaf Hopper Jassids 0.60 0.68 0.64 225
Leaf Reddening 0.83 0.84 0.84 578
Leaf Variegation 0.75 0.91 0.82 116
accuracy 0.75 2137
macro avg 0.74 0.76 0.74 2137
weighted avg 0.78 0.75 0.75 2137
(Continued)
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org16

respectively, whereas ShuffleNetV2 has higher error rates of 40.40%
for Bacterial Blight and 40.84% for Curl Virus.
5.9 Multi-seed evaluation framework and
performance stability
To ensure the statistical rigor, reproducibility, and reliability of
the comparative evaluation, all deep learning architectures were
trained and tested across five independent random initializations
(Seeds 42 through 46). Rather than relying on a single training
outcome, model performance is quanti fied and reported as Mean
Accuracy ± Standard Deviation (SD). This multi-run evaluation
accounts for training stochasticity and weight initialization vari-
ance. As detailed in Table 3 ,E f ficientNetB0 achieved the highest
overall performance, yielding a mean accuracy of 91.29% ± 0.54%,
indicating both superior classi fication capability and high training
stability due to its minimal standard deviation. Similarly, ResNet50
and ResNet101 maintained strong baseline accuracies of
90.48% ± 0.93% and 89.55% ± 0.99%, respectively. In contrast,
a r c h i t e c t u r e ss u c ha sV G G 1 6e x h i b i t e das i g n ificantly higher
standard deviation (61.88% ± 2.34%), re flecting greater sensitivity
to random seed variations.
These quantitative findings are further corroborated by the error
bar visualizations in Figure 15.T h efigure illustrates the comparative
performance and stability of the ten evaluated architectures, sorted in
ascending order by their mean classi fication accuracy along the
horizontal axis, providing a clear visual progression from baseline
frameworks to top-performing models. Each vertical column repre-
sents the mean accuracy percentage annotated directly above the bar,
with error caps denoting the standard deviation across the five runs.
As depicted, intermediate archi tectures such as MobileNetV2
(87.59%), ResNet152 (89.38%), Xception (83.09%), and
InceptionV3 (78.88%) form a robust middle tier, while
ShuffleNetV2 records a mean accuracy of 73.30%. Conversely,
deeper classical backbones like VGG16 and VGG19 record lower
mean scores (61.88% and 57.63%, respectively) alongside visibly
wider error bars, reinforcing their heightened sensitivity to random
weight initialization and training volatility.
5.10 Computational ef ficiency and
inference latency
Beyond class accuracy, a more realistic view of the scalability of
DL models requires examining inference speed and computational
complexity with regard to potential deployments. Figure 16 shows
the total GFLOPs in relation to the average inference latency (ms/
image) of each convolutional neural network tested in the compar-
ison. Deep learning models such as Xception, ResNet101, and
ResNet152 offer accurate classi fication results, but have an order
of magnitude larger average inference latencies of between 101 and
110 ms/image, due to their high computational complexity.
Residual networks offer better average inference speed than VGG,
with Xception having the fastest average at 28 ms/image, and
VGG19 only slower (33 ms/image ). More lightweight models
such as MobileNetV2 and Shuf fleNetV2 offer signi ficantly lighter
total GFLOPs (less than 2 GFLOPs) for much lower average
inference latencies of 77 –78 ms/image, and are still able to run on
resource-constrained edge devices and mobile platforms that could
be necessary in an agricultural setting. Similarly, the optimized
CNN EfficientNetB0 offers a good trade-off between high classi fi-
cation accuracy and lower complexity (1.3 GFLOPs), with an
intermediate average inference latency of 87 ms/image.
5.11 Processing speed and GFLOPs trade-
offs
In the context of model evaluation, throughput (inferences per
second) offers a complementary perspective on classi fier efficiency
under limited GPU computing resources, as shown in the total
GFLOPs versus throughput evaluation in Figure 17 . Lightweight
models such as MobileNetV2 and Shuf fleNetV2 offer the fastest
processing speeds, reaching around 12.8 to 12.9 inferences per
second with the lightest GFLOP footprints, below 2 GFLOPs.
InceptionV3, while not quite as high in throughput at around
12.2 inferences per second, remains competitive. Ef ficientNetB0
maintains high classi fication accuracy with good throughput of
around 11.4 inferences per second, while ResNet50, Xception,
TABLE 1 Continued
Model Classi fication report Precision Recall f1-score Support
EfficientNetB0 Bacterial Blight 0.90 0.82 0.86 250
Curl Virus 0.95 0.89 0.92 431
Healthy Leaf 0.86 0.99 0.92 257
Herbicide Growth Damage 1.00 0.96 0.98 280
Leaf Hopper Jassids 0.70 0.97 0.81 225
Leaf Reddening 0.98 0.85 0.91 578
Leaf Variegation 0.93 1.00 0.96 116
accuracy 0.91 2137
macro avg 0.90 0.93 0.91 2137
weighted avg 0.92 0.91 0.91 2137
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org17

TABLE 2 Comparative classi fication performance and class-wise error rates across deep learning architectures.
Model ID Class Total samples True classi fied Misclassi fications Error rate (%)
VGG16 0 Bacterial Blight 250 165 85 34.00
1 Curl Virus 431 236 195 45.24
2 Healthy Leaf 257 144 113 43.97
3 Herbicide Growth
Damage
280 273 7 2.50
4 Leaf Hopper Jassids 225 137 88 39.11
5 Leaf Reddening 578 294 284 49.13
6 Leaf Variegation 116 110 6 5.17
VGG19 0 Bacterial Blight 250 178 72 28.80
1 Curl Virus 431 251 180 41.76
2 Healthy Leaf 257 103 154 59.92
3 Herbicide Growth
Damage
280 265 15 5.36
4 Leaf Hopper Jassids 225 100 125 55.56
5 Leaf Reddening 578 250 328 56.75
6 Leaf Variegation 116 109 7 6.03
InceptionV3 0 Bacterial Blight 250 187 63 25.20
1 Curl Virus 431 369 62 14.38
2 Healthy Leaf 257 206 51 19.84
3 Herbicide Growth
Damage
280 253 27 9.64
4 Leaf Hopper Jassids 225 198 27 12.00
5 Leaf Reddening 578 378 200 34.60
6 Leaf Variegation 116 109 7 6.03
Xception 0 Bacterial Blight 250 181 69 27.60
1 Curl Virus 431 372 59 13.69
2 Healthy Leaf 257 221 36 14.01
3 Herbicide Growth
Damage
280 268 12 4.29
4 Leaf Hopper Jassids 225 204 21 9.33
5 Leaf Reddening 578 427 151 26.12
6 Leaf Variegation 116 101 15 12.93
ResNet50 0 Bacterial Blight 250 197 53 21.20
1 Curl Virus 431 364 67 15.55
2 Healthy Leaf 257 251 6 2.33
3 Herbicide Growth
Damage
280 275 5 1.79
4 Leaf Hopper Jassids 225 210 15 6.67
5 Leaf Reddening 578 519 59 10.21
6 Leaf Variegation 116 107 9 7.76
MobileNetV2 0 Bacterial Blight 250 189 61 24.40
1 Curl Virus 431 391 40 9.28
2 Healthy Leaf 257 249 8 3.11
3 Herbicide Growth
Damage
280 274 6 2.14
4 Leaf Hopper Jassids 225 205 20 8.89
5 Leaf Reddening 578 422 156 26.99
(Continued)
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org18

ResNet101, and ResNet152 show less ef ficient throughput values
between 9.0 and 9.9 inferences per second, due to increased model
complexity, with much larger numbers of parameters and the
associated GFLOP-pe r-inference overhea d. Meanwhile, VGG
models vary in performance, with higher throughput seen in the
larger-GFLOP full VGG19 model. Overall, the higher throughput
outputs underscore the fast scalability of lightweight and scaled
model structures for high-frequency agricultural monitoring.
5.12 Statistical signi ficance analysis
The performance differences between the different architectures
tested might simply be the result of random initialization error and
not a true performance difference, so we ran pairwise Student ’s t-
tests using the reference Ef ficientNetB0 model. The best Mean
accuracy of 91.29% was obtained using the architecture
EfficientNetB0 which outperforms all the others evaluated as
detailed in Table 4 . The statistical analysis shows that the mean
accuracy of ResNet50 is competitive with Ef ficientNetB0, but not
significantly different ( p = 0.1907 > 0.05). Conversely, deeper
residual variants (ResNet101 at 89.55% and ResNet152 at 89.38%)
as well as lightweight and legacy architectures exhibit a statistically
significant decline in performance ( p < 0.05) relative to the baseline.
Notably, older or heavily constrained models such as VGG19
(57.63%) and VGG16 (61.88%) demonstrated the most severe
and statistically signi ficant drops in accuracy ( p < 0.001). These
findings con firm that Ef ficientNetB0 provides an optimal balance
of parameter ef ficiency and predictive power without suffering
a statistically meaningful penalty compared to heavier resid-
ual networks.
TABLE 2 Continued
Model ID Class Total samples True classi fied Misclassi fications Error rate (%)
6 Leaf Variegation 116 107 9 7.76
ResNet101 0 Bacterial Blight 250 207 43 17.20
1 Curl Virus 431 368 63 14.62
2 Healthy Leaf 257 251 6 2.33
3 Herbicide Growth
Damage
280 270 10 3.57
4 Leaf Hopper Jassids 225 201 24 10.67
5 Leaf Reddening 578 509 69 11.94
6 Leaf Variegation 116 111 5 4.31
ResNet152 0 Bacterial Blight 250 191 59 23.60
1 Curl Virus 431 365 66 15.31
2 Healthy Leaf 257 246 11 4.28
3 Herbicide Growth
Damage
280 271 9 3.21
4 Leaf Hopper Jassids 225 209 16 7.11
5 Leaf Reddening 578 508 70 12.11
6 Leaf Variegation 116 107 9 7.76
ShuffleNetV2 0 Bacterial Blight 250 149 101 40.40
1 Curl Virus 431 255 176 40.84
2 Healthy Leaf 257 195 62 24.12
3 Herbicide Growth
Damage
280 270 10 3.57
4 Leaf Hopper Jassids 225 152 73 32.44
5 Leaf Reddening 578 485 93 16.09
6 Leaf Variegation 116 105 11 9.48
EfficientNetB0 0 Bacterial Blight 250 205 45 18.00
1 Curl Virus 431 383 48 11.14
2 Healthy Leaf 257 254 3 1.17
3 Herbicide Growth
Damage
280 270 10 3.57
4 Leaf Hopper Jassids 225 218 7 3.11
5 Leaf Reddening 578 492 86 14.88
6 Leaf Variegation 116 116 0 0.00
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org19

5.13 Model calibration and expected
calibration error analysis
Expected Calibration Error is a quantitative metric that mea-
sures the discrepancy between a model ’s predicted con fidence
scores and its empirical (true) accuracy. It groups model predictions
into discrete con fidence bins and computes a weighted absolute
difference between accuracy and con fidence within each bin. It
provides a single summary statistic to evaluate how well-calibrated a
classifier’s probability outputs are, addressing the common over-
confidence issue in deep neural networks. Mathematically ECE is
defined in Equation 8:
ECE = o
M
m=1
Bmjj
N   acc(Bm) − conf (Bm)jj (8)
where M is the total number of con fidence bins, N is the total
number of samples in the evaluation set, Bm is the subset of samples
whose predicted con fidence falls into the m-th bin, acc( Bm) is the
actual empirical accuracy of the samples in bin Bm, and conf( Bm)i s
the average predicted con fidence of the samples in bin Bm.
A reliability diagram is a visual diagnostic tool that plots
empirical accuracy (on the vertical axis) against predicted con fi-
dence (on the horizontal axis) across multiple evaluation bins.
Temperature scaling is a lightweight, post-hoc calibration tech-
nique that softens or sharpens the output probability distribution of
a trained neural network without altering its predicted class
rankings or validation accuracy. Mathematically the calibrated
probability for class i is given by Equation 9:
^P i = exp (zi=T)
oj exp (zj=T) (9)
where zi is the raw logit score for class i, T > 0 is the learned
temperature parameter, and if T > 1, the resulting probability
distribution is smoothed, reducing overcon fidence. Finally, if
T = 1, the original softmax probabilities are recovered.
To assess the reliability and probabilistic con fidence of the
evaluated deep learning architectures, we computed the Expected
TABLE 3 Multi-seed performance comparison of transfer learning architectures for cotton leaf health condition classi fication (Mean Accuracy ± SD
across five independent runs).
Model architecture Seed 42 Seed 43 Seed 44 Seed 45 Seed 46 Mean ± SD
VGG16 63.59 64.58 59.38 59.85 62.00 61.88 ± 2.27
VGG19 58.77 57.70 59.15 55.83 56.72 57.63 ± 1.39
InceptionV3 79.55 80.81 76.28 80.21 77.54 78.88 ± 1.90
Xception 83.01 83.06 84.84 83.01 81.52 83.09 ± 1.18
ResNet50 89.99 91.44 89.89 89.56 91.53 90.48 ± 0.93
MobileNetV2 85.96 88.86 87.60 87.83 87.69 87.59 ± 1.04
ResNet101 89.71 90.64 89.75 89.71 87.93 89.55 ± 0.99
ResNet152 88.77 90.45 90.03 89.66 87.97 89.38 ± 1.00
ShuffleNetV2 75.39 73.94 74.17 73.14 69.86 73.30 ± 2.09
EfficientNetB0 90.69 91.25 91.02 92.14 91.34 91.29 ± 0.54
FIGURE 15
Multi-seed performance comparison of deep learning architectures, showing mean accuracy and standard deviation (± SD) across five independent
training runs (Seeds 42 –46).
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org20

Calibration Error (ECE) both prior to and after post-hoc temper-
ature scaling. As detailed in Table 5 and illustrated in Figure 18 ,
lightweight and deep residual networks such as MobileNetV2 and
ResNet101 demonstrated exceptional baseline calibration, achieving
uncalibrated ECE values of 0.0117 and 0.0111, respectively.
However, the differences in initial calibration were higher for
deeper or structurally different architectures such as VGG16
(0.1693) or VGG19 (0.1426). The results from the post-hoc tem-
perature scaling are successful in optimizing the probability distri-
butions for all the networks tested and reach to optimal temperature
parameters (T) consistently within the range of 3.45 to 3.89 in all the
tested backbones as seen in our Figure 19. In addition, mapping the
step time of the inference versus the initial error in calibration
(shown in Figure 20), reveals that high-performing models, such as
EfficientNetB0 and MobileNetV2, continue to have fast inference
speed while also having very low initial calibration error in
comparison to heavier and legacy variants of the same backbone,
which serve as a baseline for uncertainty quantifi cation in agricul-
tural image classifi cation tasks.
Reliability diagrams or calibration curves (i.e. relationship
between model con fidence and empirical accuracy on a per-con fi-
dence bin basis) were created for all ten architectures by plotting the
mean accuracy scores of the con fidence bins shown versus the value
of the corresponding con fidence bin shown in parentheses, seen in
Figures 21a–j. The higher number of overcon fidence points and the
greater expected calibration errors (ECE) are apparent from the
pronounced deviations from the diagonal reference line of perfect
calibration by the legacies and sequential backbones (VGG16 and
VGG19) with their unconstrained fully connected parameters
(Figure 21a , b). Lightweight architectures, on the other hand,
such as Shuf fleNetV2 and Xception are closer to tracking the
diagonal, but there are some minor deviations near the extreme
probability boundaries ( Figure 21c , d). In addition, deep residual
and multi-scale networks like ResNet152, InceptionV3,
EfficientNetB0 and ResNet50 show better alignment with the
ideal diagonal reference ( Figures 21e –h). Last, ResNet101 and
MobileNetV2 result in the most stable calibration trajectory, with
the accuracies observed in their con fidence curves closely tracking
FIGURE 16
Relationship between total GFLOPs and average inference latency (ms/image) across the benchmarked models.
FIGURE 17
System throughput in inferences per second versus total GFLOPs for the benchmarked models.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org21

the ideal calibrated diagonal for most confidence bins (Figure 21i, j),
confirming the importance of a reliable and well-calibrated con fi-
dence curve using skip connections and mobile in-
verted bottlenecks.
5.14 Limitations
Although this study includes a comprehensive multi-architec-
ture benchmark, generalizability should be discussed with appro-
priate caution. First, the SAR-CLD-2024 dataset originates from a
geographically limited range of cotton field conditions, which may
restrict direct applicability to cultivars grown under substantially
different edaphic, climatic, or agricultural management practices.
Furthermore, the lack of independent external validation from
cross-regional or multi-center field datasets restricts immediate
conclusions regarding broader geographic robustness, and inde-
pendent validation across geographically distinct field datasets is
necessary before strong real-world diagnostic claims can be estab-
lished. Second, the seven-class disease taxonomy assumes mutually
exclusive categories and does not re flect common real-world field
situations where multiple pathogens are simultaneously present on
the same leaf. Third, while parameter counts indicate that heavy
architectures such as ResNet-101, ResNet-152, and Xception
demand high computational resource usage —preventing direct
edge deployment without post-training model compression or
quantization—actual empirical physical edge-device benchmarking
(e.g., latency and energy pro filing on embedded hardware) was not
performed. Finally, the models remain potentially susceptible to
learning dataset-speci fic visual shortcuts or background artifacts
rather than purely invariant pathological features. Future research
will focus on external validation using geographically diverse field
datasets, lightweight distillation approaches, uncertainty quanti fi-
cation trials, and multi-modal data integration to further improve
the robustness of the diagnostic process.
6 Conclusions
In this paper, we have systematically benchmarked automated
multi-class cotton leaf health condition classi fication using ten
state-of-the-art pretrained CNN architectures on the SAR-CLD-
2024 dataset with a common fine-tuning procedure. To prevent
majority-class bias during optimization, we utilized an augmented
training partition curated to 7,000 balanced images, while retaining
the original test partition of 2,137 images across seven categories in
its natural, imbalanced distribution to evaluate real-world opera-
tional performance. For fine-grained classi fication of agricultural
TABLE 4 Statistical signi ficance analysis and pairwise t-test results comparing various architectures against the Ef ficientNetB0 baseline.
Model architecture Mean accuracy (%) t-statistic p-value Significant vs.EfficientNetB0
EfficientNetB0 (baseline) 91.288 0 1 Baseline (Reference)
ResNet50 90.482 1.573553 0.190702 No ( p ≥ 0.05)
ResNet101 89.548 3.367265 0.028115 Yes ( p < 0.05)
ResNet152 89.376 4.015230 0.015927 Yes ( p < 0.05)
MobileNetV2 87.588 9.206943 0.000773 Yes ( p < 0.05)
Xception 83.088 13.096803 0.000196 Yes ( p < 0.05)
InceptionV3 78.878 15.342516 0.000105 Yes ( p < 0.05)
ShuffleNetV2 73.300 17.074115 0.000069 Yes ( p < 0.05)
VGG16 61.880 25.735236 0.000014 Yes ( p < 0.05)
VGG19 57.634 39.957231 0.000002 Yes ( p < 0.05)
TABLE 5 Expected calibration error (ECE), optimal temperature parameters ( T), and inference step times across evaluated deep learning models.
Deep learning
model Uncalibrated ECE Optimal temperature ( T) Calibrated ECE Inference time
MobileNetV2 0.0117 3.8483 0.3769 33 ms/Image
ResNet101 0.0111 3.8694 0.3799 33 ms/Image
ResNet50 0.0137 3.8599 0.3970 31 ms/Image
EfficientNetB0 0.0264 3.8902 0.4217 31 ms/Image
InceptionV3 0.0288 3.7570 0.4281 32 ms/Image
ResNet152 0.0359 3.7567 0.4331 35 ms/Image
Xception 0.0365 3.7922 0.4503 66 ms/Image
ShuffleNetV2 0.0392 3.7762 0.3476 35 ms/Image
VGG19 0.1426 3.5638 0.3596 32 ms/Image
VGG16 0.1693 3.4566 0.4002 28 ms/Image
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org22

FIGURE 18
Uncalibrated and calibrated expected calibration error (ECE) across the evaluated deep learning models.
FIGURE 19
Learned optimal temperature parameter ( T) across architectures.
FIGURE 20
Inference time versus baseline expected calibration error across models.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org23

diseases, Ef ficientNetB0 achieved the highest mean accuracy of
91.29% across repeated runs, proving to be an optimal accuracy –
parameter trade-off by combining compound scaling with mobile
inverted bottleneck convolutions. Pairwise t-test comparisons con-
firm that while EfficientNetB0 outperforms ResNet50 (90.48%), this
margin is not statistically signifi cant (p ≥ 0.05), indicating that both
architectures provide highly competitive top-tier performance. Both
ResNet-50 and ResNet-101 achieved strong macro F1-scores
around 0.90, reinforcing the robust and consistent effectiveness of
residual learning architectures in this multi-class environment,
though deeper variants like ResNet101 and ResNet152 exhibited a
statistically signi ficant performance decline relative to the
EfficientNetB0 baseline ( p < 0.05). MobileNetV2 achieved a com-
petitive mean accuracy of 87.59% with a much smaller computa-
tional burden, making it an ideal architecture for on-device edge
deployment. Conversely, VGG series architectures underperformed
FIGURE 21
Reliability diagrams (calibration curves) of the evaluated models: (a) VGG16, (b) VGG19, (c) ShuffleNetV2, (d) Xception, (e) ResNet152, (f) InceptionV3,
(g) EfficientNetB0, (h) ResNet50, (i) ResNet101, and (j) MobileNetV2.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org24

significantly relative to the benchmark median, with VGG16 and
VGG19 showing statistically signi ficant drops in accuracy
(p < 0.001), demonstrating that their heavy, fully connected
parameters are unsuited for the minute distinctions required in
cotton disease classi fication. Ef ficientNetB0 attains approximately
18% higher accuracy on the cotton disease test set compared to the
ShuffleNetV2 baseline, and over 33% higher accuracy than VGG19.
These findings, supported by rigorous multi-seed uncertainty esti-
mates and hypothesis testing, indicate that ResNet-50 and
EfficientNetB0 are the preferred models for developing deploy-
ment-ready cotton leaf health condition monitoring systems, while
MobileNetV2 remains best suited for resource-constrained edge
scenarios. Moving forward, this benchmark should be expanded to
include more cotton varieties, combined with hybrid architectures
integrating attention mechanisms and vision transformers, and
tested on future field data to continually advance deep learning-
driven agricultural diagnostics.
Data availability statement
The SAR-CLD-2024 dataset analyzed in this study is publicly
available in Mendeley Data ( Bishshash et al., 2024 )a t https://data.
mendeley.com/datasets/b3jy2p6k8w/2 . Further inquiries can be
directed to the corresponding authors.
Author contributions
Q-u-aA: Conceptualization, Investigation, Methodology, Software,
Writing– original draft. JX: Conceptualization, Data curation, Funding
acquisition, Project administration, Resources, Writing – review &
editing. WM: Conceptualization, Investigation, Methodology, Software,
Validation, Writing – original draft. MM: Conceptualization,
Methodology, Software, Supervision, Validation, Writing– review &
editing, Funding ac quisition. AK: Conceptualization, Data curation,
Formal analysis, Visualization, Writing – review & editing. AL:
Conceptualization, Data curation, Formal analysis, Investigation,
Visualization, Writing– review & editing.
Funding
The author(s) declared that financial support was received for
this work and/or its publication. National Social Science
Foundation General Project: Research on Promotion Strategies
and Ef ficiency Enhancement of Loc al Government Ecological
Governance (No. 23BZZ043). Shanxi Provincial Philosophy and
Social Sciences Planning Project: Research on Policy Coordination
and Effect Enhancement for Ecological Protection and High-
Quality Development in the Yellow River Basin (No.: 2025YB054).
Conflict of interest
The author(s) declared that this work was conducted in the
absence of any commercial or financial relationships that could be
construed as a potential con flict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the
creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this
article has been generated by Frontiers with the support of arti ficial
intelligence and reasonable e fforts have been made to ensure
accuracy, including review by the authors wherever possible. If
you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors
and do not necessarily represent those of their af filiated organiza-
tions, or those of the publisher, the editors and the reviewers. Any
product that may be evaluated in this article, or claim that may be
made by its manufacturer, is not guaranteed or endorsed by
the publisher.
References
Abayeva, S., and Kabdrakhmanova, S. (2026). Digital technologies in crop production:
A scoping review with transferability analysis for Central Asia. AgriEngineering 8, 199.
doi: 10.3390/agriengineering8050199
Ahmad, A. M., Abdullah, M., Moon, H., and Han, D. (2021). Plant disease detection in
imbalanced datasets using ef ficient convolutional neural networks with stepwise
transfer learning. IEEE Access 9, 140565 –140580. doi: 10.1109/access.2021.3119655
Ahmed, M. (2021). Leveraging convolutional neural network and transfer learning for
cotton plant and leaf disease recognition. Int. J. Image Graphics Signal. Process. 13, 47.
doi: 10.5815/ijigsp.2021.04.04
Ali, B., Talib, R., Hanif, M. K., and Sarwar, M. U. (2026). A hybrid deep learning
approach for plant disease classi fication: Integrating Ef ficientNet and vision trans-
formers. J. Anim. Plant Sci. 36, 79 –96. doi: 10.36899/JAPS.2026.1.0007
Anwar, S., Soomro, S. R., Baloch, S. K., Patoli, A. A., and Kolachi, A. R. (2023).
Performance analysis of deep transfer learning models for the automated detection of
cotton plant diseases. E n g i n e e r i n gT e c h n o l .A p p l .S c i .R e s .13, 11561 –11567.
doi: 10.48084/etasr.6187
Arnal Barbedo, J. G. (2013). Digital image processing techniques for detecting,
quantifying and classifying plant diseases. SpringerPlus 2, 660. doi: 10.1186/2193-
1801-2-660
Aslam, A., Usman, S. M., Zubair, M., Yasin, A., Owais, M., and Hussain, I. (2025).
Multi-convolutional neural networks for cotton disease detection using synergistic
deep learning paradigm. PloS One 20, e0324293. doi: 10.1371/journal.pone.0324293
Attallah, O. (2023). Tomato leaf disease classi fication via compact convolutional neural
networks with transfer learning and feature selection. Horticulturae 9, 149.
doi: 10.3390/horticulturae9020149
Bishshash, P., Nirob, A. S., Shikder, H., Sarower, A. H., Bhuiyan, T., and Noori, S. R. H.
(2024). A comprehensive cotton leaf disease dataset for enhanced detection and
classification. Data Brief. 57, 110913. doi: 10.1016/j.dib.2024.110913
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org25

Chaube, H., and Pundhir, V. (2005). Crop Diseases and Their Management (Prentice-
Hall of India Private Limited).
Chavan, K. A., and Shirdhonkar, M. (2024). “Real time cotton crop disease detection
using deep transfer learning ”, in: 2024 second international conference on advances in
information technology (ICAIT) (Chikkamagaluru, Karnataka: IEEE). 9. 1 –5.
doi: 10.1109/ICAIT61638.2024.10690393
Chellamani, G. K., N, A., Manoj, A., Hariharan, V., Palaniyappan, S., and SK, S. (2026).
A modular vision transformer and machine learning framework for scalable multi-crop
plant disease classi fication. Sci. Rep . doi: 10.1038/s41598-026-65004-7
Chohan, S., Perveen, R., Abid, M., Tahir, M. N., and Sajid, M. (2020). “Cotton diseases
and their management, ” in Cotton Production and Uses: Agronomy, Crop Protection,
and Postharvest Technologies (Singapore: Springer), 239 –270.
Chollet, F. (2017). “Xception: Deep learning with depthwise separable convolutions ”,
in: 2017 IEEE conference on computer vision and pattern recognition (CVPR)
(Honolulu, HI: IEEE). p. 1251 –1258. doi: 10.1109/CVPR.2017.195
Cui, Y., Song, Y., Sun, C., Howard, A., and Belongie, S. (2018). “Large scale fine-grained
categorization and domain-specific transfer learning”, in: 2018 IEEE/CVF conference on
computer vision and pattern recognition (Salt Lake City, UT: IEEE). doi: 10.1109/
CVPR.2018.00432
Dhavale, M. C., Aiwale, H. N., and Sarnobat, D. H. (2025). Modern diagnostic techniques
in plant pathology. Plant Health Dis. Management: A. Modern Approach,1 8 4 .
Dong, X., Wang, Q., Huang, Q., Ge, Q., Zhao, K., Wu, X., et al. (2023). PDDD-
PreTrain: A series of commonly used pre-trained models support image-based plant
disease diagnosis. Plant Phenomics 5, 54. doi: 10.34133/plantphenomics.0054
Ferentinos, K. P. (2018). Deep learning models for plant disease detection and
diagnosis. Comput. Electron. Agric. 145, 311 –318. doi: 10.1016/j.compag.2018.01.009
Gholizade, M., Soltanizadeh, H., Rahmanimanesh, M., and Sana, S. S. (2025). A review
of recent advances and strategies in transfer learning. Int. J. System Assur. Eng. Manage.
16, 1123 –1162. doi: 10.1007/s13198-024-02684-2
Ghosal, A. (2022). “Management guide of sucking insect pest, ” in Insecticides in Pest
Control—Impact, Challenges and Strategies (IntechOpen).
Guo, Y. Y., Cai, H., Bramley, G. L., Atkins, H. J., Li, B., and Theodossiades, S. (2025). A
review of arti ficial intelligence in herbarium specimen image analysis. Artif. Intell. Rev.
58, 402. doi: 10.1007/s10462-025-11408-2
Haqiq, N., et al. (2026). Lightweight hybrid deep learning for strawberry disease
recognition and edge deployment using dynamic multi-scale CNN –transformer fusion.
AgriEngineering 8, 75. doi: 10.3390/agriengineering8020075
Haque, M. E., Saykat, M. T. H., Al-Imran, M., Siam, A. H., Uddin, J., and Ghose, D.
(2026). An attention enhanced CNN ensemble for interpretable and accurate cotton
leaf disease classi fication. Sci. Rep. 16, 4476. doi: 10.1038/s41598-025-34713-w
He, K., Zhang, X., Ren, S., and Sun, J. (2016). “Deep residual learning for image
recognition”, in: 2016 IEEE Conference on Computer Vision and Pattern Recognition
(CVPR). (Las Vegas, NV). p. 770 –778. doi: 10.1109/CVPR.2016.90
Hossain, M. M., Sultana, F., Mostafa, M., Ferdus, H., Rahman, M., Rana, J. A., et al.
(2024). Plant disease dynamics in a changing climate: Impacts, molecular mechanisms,
and climate-informed strategies for sustainable management. Discover Agric. 2, 132.
doi: 10.1007/s44279-024-00144-w
Islam, M. M., Talukder, M. A., Sarker, M. R. A., Uddin, M. A., Akhter, A., Sharmin, S.,
et al. (2023). A deep learning model for cotton disease prediction using fine-tuning with
smart web application in agriculture. Intell. Syst. Appl. 20, 200278. doi: 10.1016/
j.iswa.2023.200278
Jafar, A., Bibi, N., Naqvi, R. A., Sadeghi-Niaraki, A., and Jeong, D. (2024). Revolutionizing
agriculture with artificial intelligence: Plant disease detection methods, applications, and
their limitations. Front. Plant Sci. 15, 1356260. doi: 10.3389/fpls.2024.1356260
Jawed, M. M., Tufail, F. A., Ahmed, M. Z., R, A. S., Nallusamy, P., and Raja, K. T.
(2026). A hybrid deep learning framework using convolutional and transformer models
for robust plant disease classi fication. Sci. Rep. 16, 9704. doi: 10.1038/s41598-026-
38209-z
Johri, P., Kim, S., Dixit, K., Sharma, P., Kakkar, B., Kumar, Y., et al. (2024). Advanced
deep transfer learning techniques for ef ficient detection of cotton plant diseases. Front.
Plant Sci. 15, 1441117. doi: 10.3389/fpls.2024.1441117
Junaidi, A., Hashim, S. Z. M., Othman, M. S. B., Mohamad, M. M. B., Alhussian, H.,
Abdulkadir, S. J., et al. (2025). Deep learning and edge computing in agriculture: A
comprehensive review of recent trends and innovations. IEEE Access 13, 137464 –
137490. doi: 10.1109/access.2025.3595621
Kamilaris, A., and Prenafeta-Boldú, F. X. (2018). Deep learning in agriculture: A survey.
Comput. Electron. Agric. 147, 70–90. doi: 10.1016/j.compag.2018.02.016
Kaur, G., Al-Yarimi, F. A. M., Bharany, S., Rehman, A. U., and Hussen, S. (2025).
Explainable AI for cotton leaf disease classi fication: A metaheuristic-optimized deep
learning approach. Food. Sci. Nutr. 13, e70658. doi: 10.1002/fsn3.70658
Kumar, R. V., and Thirupati, D. G. (2026). DenseNet169-enabled high-accuracy
automated detection system for cotton leaf diseases using deep transfer learning. Int.
J. Science Strategic Manage. Technol. 02, 1–9. doi: 10.55041/ijsmt.v2i8.064
LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature 521, 436 –444.
doi: 10.1109/hotchips.2015.7477328
Ma, N., Zhang, X., Zheng, H. T., and Sun, J. (2018). “ShuffleNet V2: Practical guidelines
for ef ficient CNN architecture design ”, in: European conference on computer vision
(Munich: Springer). p. 122 –138.
Mazumder, M. K. A., Mridha, M. F., Alfarhood, S., Safran, M., Abdullah-Al-Jubair, M.,
and Che, D. (2024). A robust and light-weight transfer learning-based architecture for
accurate detection of leaf diseases across multiple plants using less amount of images.
Front. Plant Sci. 14. doi: 10.3389/fpls.2023.1321877
Mudafale, P. (2026). Intelligent cotton leaf disease detection system using deep learning
and multi-layer validation framework. Int. J. For. Res. Appl. Sci. Eng. Technol. 14, 1705–
1711. doi: 10.22214/ijraset.2026.79486
Mulugu, N., Praharika, A., Reddy, A. R., Sahitya, A., Hasini, A., and Abhignya, C.
(2025). “High-accuracy cotton leaf disease classi fication using optimized DenseNet121
with transfer learning and fine-tuning”, in: 2025 IEEE international conference on
communication networks and computing (CNC) (Sonbhadra: IEEE). p. 862 –868.
doi: 10.1109/CNC68716.2025.11484724
Nalwanga, R., Spreeuwers, L., Talavera, E., and Owomugisha, G. (2025). Multi-crop
disease detection in computer vision for resource-constrained farms —a review. IEEE
Access 13, 217265 –217284. doi: 10.1109/access.2025.3647114
Ngugi, H. N., Ezugwu, A. E., Akinyelu, A. A., and Abualigah, L. (2024). Revolutionizing
crop disease detection with computational deep learning: A comprehensive review.
Environ. Monit. Assess. 196, 302. doi: 10.1007/s10661-024-12454-z
Nivethitha, T., and Venkataramanan, C. (2026). A quantum-enhanced deep learning
framework for precise leaf disease detection in tomato and cotton crops. Iranian J. Sci.
Technology Trans. Electrical Eng. 50, 323 –354. doi: 10.1007/s40998-025-00912-4
Patra, A. K., and Gajurel, T. (2025). “Improved cotton leaf disease classi fication using
parameter-efficient deep learning framework ”, in: 2025 joint international conference
on digital arts, media and technology with ECTI northern section conference on
electrical, electronics, computer and telecommunications engineering (ECTI DAMT &
NCON) (Nan: IEEE). p. 737 –742. doi: 10.1109/ECTIDAMTNCON64748.
2025.10961983
Pethybridge, S. J., and Nelson, S. C. (2015). Leaf Doctor: A new portable application for
quantifying plant disease severity. Plant Dis. 99, 1310 –1316. doi: 10.1094/pdis-03-15-
0319-re
Potharaju, S., et al. (2026). An explainable vision transformer model with transfer
learning for accurate bean leaf disease classi fication. Sci. Rep. 16, 10402. doi: 10.1038/
s41598-026-41723-9
Pujari, J. D., Yakkundimath, R., and Byadgi, A. S. (2016). SVM and ANN based
classification of plant diseases using feature reduction technique. Int. J. Interact.
Multimedia Artif. Intell. 3, 6 –14. doi: 10.9781/ijimai.2016.371
Rajendran, T., Sathya, S., Jenifer, J. V., Abiniya, S., Keerthana, B., and Kiran, B. (2026).
“Edge-AI driven smart crop protection and management: Integrating vision trans-
formers, graph neural networks, and federated learning for real-time, sustainable, and
scalable precision agriculture ”, in: 2026 International Conference on Machine Learning
and Autonomous Systems (ICMLAS) .( B a n g k o k ) .p .4 4 4–451. doi: 10.1109/
ICMLAS67792.2026.11483902
Sain, S. K., Monga, D., Sabesh, M., Prasad, Y. G., Prakash, A. H., and Singh, R. K.
(2023). Diseases and Disorders of Cotton: A field guide for symptom-based identi fication
and management . (Tamil Nadu: ICAR-AICRP on Cotton, ICAR-CICR, Regional
Station), 140.
Sajjan, M. M., Petimani, B. S., Sarala, D., Surasura, P., Kamatar, V., and Hadimani, B. A.
(2025). “Cotton plant diseases classi fication using deep learning ”, in: 2025 IEEE north
karnataka subsection flagship international conference (NKCon) (Hubballi: IEEE). p. 1–
6. doi: 10.1109/NKCon66957.2025.11345780
Saleem, S., Sharif, M. I., Sharif, M. I., Sajid, M. Z., and Marinello, F. (2024). Comparison
of deep learning models for multi-crop leaf disease detection with enhanced vegetative
feature isolation and de finition of a new hybrid architecture. Agronomy 14, 2230.
doi: 10.3390/agronomy14102230
Saleem, M. H., Potgieter, J., and Arif, K. M. (2021). Automation in agriculture by
machine and deep learning techniques: A review of recent developments. Precis. Agric.
22, 2053 –2091. doi: 10.1007/s11119-021-09806-x
S a n d l e r ,M . ,H o w a r d ,A . ,Z h u ,M . ,Z h m o g i n o v ,A . ,a n dC h e n ,L .C .( 2 0 1 8 ) .
“MobileNetV2: Inverted residuals and linear bottlenecks ”, in: 2018 IEEE/CVF confer-
ence on computer vision and pattern recognition (Salt Lake City, UT: IEEE). p. 4510 –
4520. doi: 10.1109/CVPR.2018.00474
Shafik, W., Tufail, A., De Silva Liyanage, C., and Apong, R. A. A. H. M. (2024). Using
transfer learning-based plant disease classi fication and detection for sustainable
agriculture. BMC Plant Biol. 24, 136. doi: 10.1186/s12870-024-04825-y
Shrikhande, D., and Gawade, S. (2026). An interpretable multimodal transfer learning
decision-support framework for multi-crop, multi-disease detection in precision
agriculture. Int. J. Aquat. Res. Environ. Stud. 6, 955 –967. doi: 10.70102/IJARES/V6I2/
6-2-1593
Shrivastava, V. K., Pradhan, M. K., Minz, S., and Thakur, M. P. (2019). Rice plant
disease classi fication using transfer learning of deep convolution neural network. Int.
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org26

Arch. Photogrammetry Remote Sens. Spatial Inf. Sci. 42, 631 –635. doi: 10.5194/isprs-
archives-xlii-3-w6-631-2019
Simonyan, K., and Zisserman, A. (2015). Very deep convolutional networks for large-scale
image recognition. In: 2015 IEEE 3rd IAPR Asian Conference on Pattern Recognition
(ACPR), (Kuala Lumpur), 730–734. doi: 10.48550/arXiv.1409.1556
Singh, V., and Misra, A. K. (2017). Detection of plant leaf diseases using image
segmentation and soft computing techniques. Inf. Process. Agric. 4, 41–49. doi: 10.1016/
j.inpa.2016.10.005
Szegedy, C., Vanhoucke, V., Ioffe, S. ,S h l e n s ,J . ,a n dW o j n a ,Z .( 2 0 1 6 ) .“Rethinking the
inception architecture for computer vision”,i n :Proceedings of the IEEE conference on computer
vision and pattern recognition(Las Vegas, NV). p. 2818–2826. doi: 10.1109/CVPR.2016.308
Tan, M., and Le, Q. (2019). “EfficientNet: Rethinking model scaling for convolutional
neural networks ”, in: International conference on machine learning (Long Beach, CA:
PMLR). p. 6105 –6114.
Tugrul, B., Elfatimi, E., and Eryigit, R. (2022). Convolutional neural networks in detection
of plant leaf diseases: A review. Agriculture 12, 1192. doi: 10.3390/agriculture12081192
Upadhyay, A., Chandel, N. S., Singh, K. P., Chakraborty, S. K., Nandede, B. M., Kumar,
M., et al. (2025). Deep learning and computer vision in plant disease detection: A
comprehensive review of techniques, models, and trends in precision agriculture. Artif.
Intell. Rev. 58, 92. doi: 10.1007/s10462-024-11100-x
Vallabhajosyula, S., Sistla, V., and Kolli, V. K. K. (2022). Transfer learning-based deep
ensemble neural network for plant leaf disease detection. J. Plant Dis. Prot. 129, 545 –
558. doi: 10.1007/s41348-021-00465-8
Verma, K., Dubey, M. K., and Pandey, D. K. (2026). Application of deep transfer
learning for foliar disease detection in solanaceous vegetables for sustainable produc-
tion. Int. J. Veg. Sci. ,1 –40. doi: 10.1080/19315260.2026.2714102
Wang, K., Guo, S., Yang, X., Lin, B., Cai, W., Tang, R., et al. (2026). An integrated edge
AI prototype for smart agriculture: Real-time pest detection, physical trapping, and
multi-node deployment analysis under field uncertainty. Front. Plant Sci. 17.
doi: 10.3389/fpls.2026.1853368
Yang, Z. Y., Xia, W. K., Chu, H. Q., Su, W. H., Wang, R. F., and Wang, H. (2025). A
comprehensive review of deep learning applications in cotton industry: From field
monitoring to smart processing. Plants 14, 1481. doi: 10.3390/plants14101481
Zhu, H., Lin, C., Liu, G., Wang, D., Qin, S., Li, A., et al. (2024). Intelligent agriculture:
Deep learning in UAV-based remote sensing imagery for crop diseases and pests
detection. Front. Plant Sci. 15, 1435016. doi: 10.3389/fpls.2024.1435016
Abro et al. 10.3389/fpls.2026.1993878
Frontiers in Plant Science frontiersin.org27

论文原文与出版信息

Frontiers in Plant Science
查看论文原文

继续探索