论文信息
论文解读
AI 辅助整理- 方法
- EfficientNet深度特征、长程注意力、自适应模糊加权与轻量ShCNN分类
- 发现
- 四类棉花叶病分类准确率达98.72%,参数量4.87M,单图推理3.79ms。
- 机会点
- 可探索模糊加权与注意力机制在田间复杂背景、多作物病害跨域泛化中的轻量化部署。
论文摘要
棉花叶片病害的早期准确识别对于提高作物生产力和支撑精准农业具有重要意义。本研究提出了一种EfficientNet-Fuzzy ShCNN架构,用于复杂成像条件下的多类棉花叶片病害分类。该框架集成了基于EfficientNet的深层特征提取、长程注意力机制、自适应模糊特征加权以及轻量级深度可分离ShCNN分类器,以提升分类性能和鲁棒性。模型采用公开可用的Kaggle棉花叶片病害数据集进行评估,该数据集包含四类:健康、细菌性枯萎病、曲叶病毒病和镰刀菌枯萎病。经过数据集筛选和实验子集构建,保留1,400张原始图像,并在数据增强前将其分层划分为训练、验证和独立测试子集。数据增强仅应用于训练子集,以防止信息泄漏。实验结果表明,所提框架实现了98.72%的分类准确率,精确率、召回率和F1分数均超过98%,错误率为1.28%。该框架还表现出计算效率,可训练参数为487万,计算量为0.46 GFLOPs,神经网络平均推理时间为每张图像3.79 ms,表明在所评估的GPU环境下具有较低的推理延迟。对比评估表明,相对于所评估的基线模型和近期深度学习架构,所提框架在分类性能与计算复杂度之间提供了良好的平衡。逐组件消融分析进一步证明了预处理、注意力、模糊加权、特征融合、ShCNN和优化各阶段的贡献。在匹配训练种子上的Friedman重复测量分析识别出所评估架构之间存在总体差异(𝜒 𝐹 2, (4) = 20.00,𝑝 ≈ 0.00050)。此外还进行了带Holm校正的精确配对Wilcoxon符号秩事后检验;然而,由于仅有五次匹配运行可用,各成对比较未达到传统的0.05显著性水平。
全文
International journal of intelligent engineering and systemsReceived: September 4, 2026. Revised: October 3, 2026. 370
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
EfficientNet–Fuzzy ShCNN Framework for Multi-class Cotton Leaf Disease
Classification Using Adaptive Feature Fusion
Sunita Chalageri1 Kavita K Patil1 Manjula H Nebagiri2 Saba Tahseen3
Naitik S T4 Sandeep Kumar5 Shrikant J. Honade6 G. J. Sahani7 Syed Sumera Ali8
Sridhar S K9*
¹Department of Information Science and Engineering, Global Academy of Technology, Bengaluru, Karnataka,
India
2Department of Computer Science and Engineering,
Shridevi Institute of Engineering and Technology, Tumkur, India
3Department of Computer Science and Design,
Dayananda Sagar College of Engineering, Bengaluru, Karnataka, India
4Department of Computer Science and Engineering, Dayananda Sagar University, Bengaluru, Karnataka, India
5Department of Computer Science and Engineering, Symbiosis Institute of Technology, Nagpur Campus,
Symbiosis International (Deemed University), Pune, Maharashtra, India
6Department of Electronics Engineering (VLSI Design & Technology),
CSMSS Chh. Shahu College of Engineering, Chhatrapati Sambhajinagar, Maharashtra, India
7Department of Computer Science and Engineering,
CSMSS Chh. Shahu College of Engineering, Chhatrapati Sambhajinagar, Maharashtra, India
8Department of Electronics and Communication (ACT),
CSMSS Chh. Shahu College of Engineering, Chhatrapati Sambhajinagar, Maharashtra, India
9Department of Computer Science and Engineering,
School of Engineering, Dayananda Sagar University, Bengaluru South, Karnataka, India
- Corresponding author’s Email: sridharsk.dsu@gmail.com
Abstract: The early and accurate recognition of cotton leaf diseases is important for improving crop productivity and
supporting precision agriculture. This study proposes an EfficientNet –Fuzzy ShCNN architecture for multi -class
cotton leaf disease classification under complex imaging conditions. The proposed framework integrates EfficientNet-
based deep feature extraction, a long -range attention mechanism, adaptive fuzzy feature weighting, and a lightweight
depthwise ShCNN classifier to improve classificati on performance and robustness. The model was evaluated using a
publicly available Kaggle Cotton Leaf Disease dataset containing four classes: Healthy, Bacterial Blight, Leaf Curl
Virus, and Fusarium Wilt. Following dataset screening and experimental -subset construction, 1,400 original images
were retained and stratified into training, validation, and independent test subsets prior to augmentation. Data
augmentation was applied exclusively to the training subset to prevent information leakage. Experimental results show
that the proposed framework achieves a classification accuracy of 98.72%, with precision, recall, and F1 -score values
exceeding 98% and an error rate of 1.28%. The framework also demonstrates computational efficiency with 4.87
million trainable parameters, 0.46 GFLOPs, and an average neural -network inference time of 3.79 ms per image,
indicating low inference latency under the evaluated GPU environment. Comparative evaluation demonstrates that the
proposed framework provides a favorable balance between classification performance and computational complexity
relative to the evaluated baseline and recent deep -learning architectures. Component -wise ablation analysis further
demonstrates the contribution of the preprocessing, attention, fuzzy weighti ng, feature -fusion, ShCNN, and
optimization stages. A Friedman repeated -measures analysis across the matched training seeds identified an overall
difference among the evaluated architectures (𝜒𝐹2, (4) = 20.00 𝑝 ≈ 0.00050). Exact paired Wilcoxon signed -rank
post-hoc tests with Holm correction were additionally performed; however, the individual pairwise comparisons did
not reach the conventional 0.05 significance level because only five matched runs were available.
Received: September 4, 2026. Revised: October 3, 2026. 371
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
Keywords: Cotton leaf disease, Attention mechanism, EfficientNet, Deep learning, Image -based disease detection,
Precision agriculture.
- Introduction
Cotton is one of the most commercially important
commercial crops in the world, being essential for the
textile sector and significantly contributing to the
global agriculture economy [1]. Cotton crop
productivity is greatly influenced by plant health,
which is adversely impacted by a number of leaf
diseases such as bacterial blight, leaf curl virus,
fusarium wilt; If left undetected and uncontrolled,
these diseases can spread rapidly over a large area
resulting in significant yield losses. Therefore, timely
and correct diagnosis of the cotton leaf disorders is
vital.
Currently, disease identification relies primarily
on manual inspection by agricultural specialists
which is not only time-consuming but also subjective
and prone to human error. Moreover, due to the
requirement of real -time monitoring, manual
diagnosis is not possible in large scale farming
scenarios. These limitations have led to carrying out
the plant disease detection widely which gives rise to
the automatic and intelligent system for plant disease
detection. Deep learning is a branch of artificial
intelligence that has rapidly advanced in the last few
years, followed by great interest in automating
disease detection from image [2]. Early methods were
largely of conventional machine learning algorithms
(e.g. Support Vector Machines (SVM), k -Nearest
Neighbours (k -NN) and Random Forest classifiers.
Although these algorithms derive the predictions
based on hand -built features (colours, textons and
shape descriptors). Although these methods led to
early improvements, their performance is limited due
to dependence on feature engineering and they fail to
capture complex patterns from high-dimensional data.
Recently, CNN models have been successfully
applied to the classification of plant diseases as they
can automatically learn hierarchical features from
material [3]. Deep convolutional neural network
(CNN) architectures, namely AlexNet, VGGNet19
and ResNet have been used in a few studies to
enhance classification precision. Since then, more
advanced models such as DenseNet and EfficientNet
have improved performance with better feature reuse
and scaling strategy. Furthermore, attention -based
CNN models have b een proposed to emphasise
disease-relevant areas, whereas transformer
architectures reflect the capacity to capture global
contextual dependencies [4].
Despite these advances, a few issues remained
unsolved. Images taken from the real world of
agriculture tend to have complex backgrounds with
inconsistent lighting conditions and noise, which
affects the performance of the model. Moreover,
typical CNN -based models focus mainly on local
feature extraction and fail to capture long -distance
dependencies.
1.1 Problem statement and objective
The Cotton leaf disease classification based on
current methods with deep learning generally
involves the key task of differentiating several classes
of diseases under specific environments related to
focused illumination, background noise, and other
similar visual features. Deep learning systems trained
on such data currently struggle to generalize across
these variations, even if they won’t overlap between
classes. Thus, the objective of this study is to develop
a reliable and effective classification fra mework
which would efficiently detect, classify cotton leaf
diseases with addressing these challenges.
Specifically, the proposed research focuses on:
• Enhance classification accuracy to identify multi-
class cotton leaf diseases.
• Enhance robustness to environmental variability
and noise.
• Introduce mechanisms for handling of
uncertainty in feature representation.
• Construct a Model that is computationally
permissible and applied in the actual world to
utilize.
1.2 Main contributions
The main contributions of this research are
outlined in the following summaries:
• A novel EfficientNet –Fuzzy ShCNN hybrid
architecture is proposed that jointly performs
deep feature extraction, long -range contextual
learning, uncertainty -aware feature refinement,
and lightweight disease classification within a
unified framework.
• An Efficient Long -Range Attention mechanism
is introduced to capture global contextual
dependencies that conventional EfficientNet
architectures cannot effectively model.
Received: September 4, 2026. Revised: October 3, 2026. 372
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
• An adaptive fuzzy feature weighting module
dynamically assigns importance to extracted
features, improving robustness against
illumination changes, background noise, and
visually similar disease symptoms.
• A lightweight depthwise ShCNN classifier
significantly reduces computational complexity
compared with transformer -based classifiers
while maintaining high classification
performance.
• A Bluefin Trevally Optimization Algorithm
(BTOA) automatically optimizes network
hyperparameters, resulting in improved
convergence and enhanced generalization.
• Extensive experiments including comparative
analysis, ablation studies, statistical significance
testing, and uncertainty estimation demonstrate
the superiority of the proposed framework.
The remainder of this paper is organized as
follows. Section 2 reviews existing approaches for
cotton leaf disease classification and discusses the
research gap and motivation. Section 3 presents the
proposed EfficientNet–Fuzzy ShCNN methodology,
including preprocessing, FLICM segmentation,
EfficientNet-based feature extraction, long -range
attention, multi-level feature fusion, adaptive fuzzy
weighting, ShCNN classification, BTOA -based
hyperparameter optimization, and implementation
details. Section 4 prese nts the experimental setup,
comparative evaluation, ablation analysis,
computational complexity, statistical validation,
uncertainty estimation, and explainability analysis.
Section 5 discusses the research implications, while
Sections 6 and 7 present the conclusions, limitations,
and directions for future work.
2. Related work
Hu et al. (2025) [5] proposed an enhanced object
detection framework named ACURS -YOLO based
on YOLOv11 for multi -class cotton leaf disease
detection. Their method incorporates a U -Net v2
module into the backbone in order to enhance the
extraction of multi -scale features. In addition, the
substitution of the SPPF module with SimSPPF
results in a reduction in the complexity of the
computations, while the C3k2_RCM module
improves the learning of long -range contextual
information respectively. The model achieve d a
mAP@0.5 of 94.6% and demonstrated superior real-
time performance (148 FPS), highlighting its
effectiveness in complex agricultural environments.
Prasad et al. (2024) [6] proposed MCCLDP, a
Multi-Class Cotton Leaf Disease Prediction
framework based on a Multi -Stream Attention -
Guided Hybrid CNN (MAH -CNN). It utilizes a
combination of deep feature extractors like
ResNet152v2 and DenseNet -121 along with a
shallow CNN to extract both global and local features.
Then a spatial attention mechanism is integrated to
highlight disease -related areas and minimize
background noise. The reported results demonstrate
strong performance for multi -class disease
classification and highlight the effectiveness of
attention-guided deep -learning architectures for
reliable disease diagnosis.
Amin et al. (2022) [7] proposed a framework for
cotton leaf disease classification using explainable
deep learning based on pre -trained VGG -16 model
[7]. They utilize coverage level data augmentation for
balancing datasets as well as multiple fully
convolutional layers to obtain anomaly maps that
point out the specific region affected by the disease.
This approach leads to transparency and trust in the
models by making incorporation of explainability. It
reached 99.99% of accuracy on real Kaggle data,
which means that it made relatively few mistakes
regarding the type of disease classified.
S. S. K. et al. (2026) [8] proposed a transformer -
based multimodal deep learning framework for early
disease detection by integrating heterogeneous
healthcare data through cross -modal fusion. Their
model improved the average detection accuracy by
7.3% and F1-score by 5.8%, highlighting the
effectiveness of transformer -based architectures for
robust feature representation and disease
classification. D et al. (2019) [9], developed a system
for processing leaf image to detect plant diseases.
This included pre-processing, extracting features and
using TensorFlow based object detection models. In
the practical use, it was operational since it predicted
a disease and even provided probable symptoms and
treatment. However, classical methods suffer from
weak general ization performance and are feature
dependent. Ahmed et al. (2026) [10] proposed a
lightweight DnCNN -based framework for efficient
image artifact suppression on resource -constrained
devices. The compressed model achieved an average
PSNR of 29.0 dB and SSIM of 0.83, while reducing
the model size by 52% and CPU inference time by
70%, demonstrating the effectiveness of lightweight
convolutional architectures for real -time edge
deployment.
2.1 Novelty and rationale of the proposed hybrid
framework
Although EfficientNet -based models have
demonstrated excellent feature extraction capability
for plant disease classification, they still exhibit
Received: September 4, 2026. Revised: October 3, 2026. 373
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
several limitations when applied to complex field
environments. EfficientNet primarily learns
hierarchical local representations but may not
sufficiently capture long -range contextual
dependencies among spatially separated disease
symptoms. To overcome the se limitations, the
proposed framework integrates five complementary
components within a unified architecture.
EfficientNet serves as an efficient backbone for
hierarchical feature extraction. An adaptive fuzzy
feature weighting mechanism estimates the
importance of extracted features under uncertain
environmental conditions, thereby suppressing noisy
or unreliable representations.
3. Methodology
The proposed methodology consists of
preprocessing, FLICM -based segmentation,
EfficientNet-based deep feature extraction, long -
range attention for capturing global spatial
dependencies, adaptive fuzzy feature weighting,
multi-level feature fusion, feature selection, and a
lightweight depthwise ShCNN classifier.
3.1 System overview
The proposed framework learns a mapping from
cotton leaf images to the corresponding disease
classes under complex imaging conditions. Let the
dataset be represented as given in Eq. (1):
𝒟 = {(𝐼𝑖, 𝑦𝑖)}𝑖=1𝑁 (1)
where, 𝐼𝑖 = input image and 𝑦𝑖 = class label. The
objective of classification is to learn a parametric
function 𝑓𝜃(⋅) and is given in Eq. (2):
𝑦̂ = 𝑓𝜃(𝐼) (2)
where, 𝑦̂ d = class -label. This entire process will
consist of data preprocessing, segmentation, deep
feature generation and attention modelling, fuzzy
weight allocation and light -weight classification
phase.
3.2 Pre-processing and normalization
To reduce the variations in illumination and pixel
intensities, we normalize each of those input images
using min–max scaling defined by Eq. (3):
𝐼𝑛𝑜𝑟𝑚 =
𝐼−𝐼𝑚𝑖𝑛
𝐼𝑚𝑎𝑥−𝐼𝑚𝑖𝑛
(3)
As shown in Eq. (3), normalization scales all
pixel values to a common range and improves
numerical stability during optimization.
Deep convolutional denoising approaches have
also been investigated for image artifact suppression
prior to feature extraction. Such preprocessing
techniques improve image quality and contribute to
more reliable feature learning under degraded
imaging cond itions [16]. A Gaussian filter is
Table 1. Qualitative Summary of Representative Cotton Leaf Disease Classification Studies
Study Year Dataset Model/ Architecture Key Techniques Optimization Accuracy (%) Explainability / Robustness
Rai et al.
[11] 2023
2293 cotton
leaf & plant
images
Improved
Deep CNN
Pooling
strategy
comparison
(max vs avg),
epoch tuning
Hyperparameter
tuning 97.98% Moderate
Arathi et
al. [12] 2023 Cotton leaf dataset DenseNet-121 (Transfer Learning) ImageNet pre-trained weights Transfer learning 91% Low
Kaur et al.
[13] 2024
Augmented
dataset
(~5000
images)
Fine-tuned
VGG16
Data
augmentation,
batch size
tuning
Hyperparameter
optimization 95.5% Moderate
Pandey et
al. [14] 2023 1710 images Hybrid Deep Learning Model Multi-disease classification Model fusion 98.9% Moderate
Rajasekar
et al. [15] 2022 Cotton dataset ResNet-50 + Xception
Deep transfer
learning +
feature
extraction
Pre-trained
optimization
98%
(validation) Moderate
Received: September 4, 2026. Revised: October 3, 2026. 374
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
Figure. 1 Proposed System Architecture
subsequently applied to suppress high -frequency
noise while preserving relevant structural
information. Then a Gaussian filter is applied to
saturate the high frequency noise in Eq. (4):
𝐼𝑠𝑚𝑜𝑜𝑡ℎ = 𝐼𝑛𝑜𝑟𝑚 ∗ 𝐺(𝜎) (4)
where, 𝐺(𝜎) = Gaussian kernel. This operation filters
noise while retaining important structural
characteristics. Rotation (±20°), horizontal flipping
(p = 0.5), brightness adjustment (±15%), and random
scaling (0.9 –1.1×) were applied exclusively to the
training subset. No augmentation or preprocessing
parameters were fitted using validation or test images.
3.3 FLICM-based leaf segmentation
FLICM Clustering is used to segment out the area
of interest that formulate the clustering objective as
shown in Eq. (5):
𝐽 = ∑ ∑ 𝑢𝑖𝑗𝑚
𝐶
𝑗=1
𝑁
𝑖=1
∥ 𝑥𝑖 − 𝑣𝑗 ∥2+ 𝜆𝐺𝑖𝑗 (5)
As presented in Eq. (5), the objective function
consists of both membership -based clustering and a
spatial constraint term 𝐺𝑖𝑗, which is described in Eq.
(6):
𝐺𝑖𝑗 = ∑ 1𝑑𝑖𝑘 + 1
𝑘∈𝑁𝑖
(1 − 𝑢𝑘𝑗)𝑚 ∥ 𝑥𝑘 − 𝑣𝑗 ∥2 (6)
The presence of the spatial term in Eq. (6)
guarantees that aggregations of the neighbouring
pixels participate in cluster assignments, which
improves segmentation robustness under noisy
environment.
3.4 EfficientNet-based deep feature extraction
The segmented image 𝐼𝑠𝑒𝑔 is forwarded through
efficientNet to extract hierarchical representations as
detailed in Eq. (7):
𝐹𝑑𝑒𝑒𝑝 = Φ𝐸𝑓𝑓𝑁𝑒𝑡(𝐼𝑠𝑒𝑔) (7)
As expressed in Eq. (7), the deep feature vector
contains high-level abstraction, including topological
and structural variances of lesions that aid in
distinguishing diseases. Fig. 2 shows the schematic
of pre -processing steps (normalized, denoised,
contrast enhanced, FLICM -based segmentation) for
cotton leaf images. Similar transformer -based
attention mechanisms have been successfully
employed in automated visual inspection applications
to capture global contextual dependencies and
improve feature representation for classification tasks
[17].
Figure. 2 Pre-processing and FLICM-Based Segmentation Pipeline for Cotton Leaf Disease Analysis
Received: September 4, 2026. Revised: October 3, 2026. 375
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
3.5 Efficient long-range attention mechanism
To capture global dependencies, an attention
mechanism is introduced. The feature transformation
into query, key, and value spaces is defined in Eq. (8):
𝑄 = 𝐹𝑑𝑒𝑒𝑝𝑊𝑄, 𝐾 = 𝐹𝑑𝑒𝑒𝑝𝑊𝐾, 𝑉 = 𝐹𝑑𝑒𝑒𝑝𝑊𝑉 (8)
Using these projections, attention weights are
computed as shown in Eq. (9):
𝐴 = 𝑆𝑜𝑓𝑡𝑚𝑎𝑥 (
𝑄𝐾𝑇
√𝑑𝑘
) 𝑉 (9)
The final refined feature representation is
obtained by combining the original and attention
features, as defined in Eq. (10):
𝐹𝑎𝑡𝑡 = 𝐹𝑑𝑒𝑒𝑝 + 𝐴 (10)
As seen in Eq. (10), residual fusion enhances
discriminative feature learning by emphasizing
globally relevant regions.
3.6 Multi-level feature extraction and fusion
To enrich representation capability, additional
features are extracted [18]. Texture features are
obtained as in Eq. (11):
𝐹𝑡𝑒𝑥𝑡𝑢𝑟𝑒 = 𝐸𝑥𝑡𝑟𝑎𝑐𝑡𝑡𝑒𝑥𝑡𝑢𝑟𝑒(𝐼𝑠𝑒𝑔) (11)
Color features are extracted using Eq. (12):
𝐹𝑐𝑜𝑙𝑜𝑟 = 𝐸𝑥𝑡𝑟𝑎𝑐𝑡𝑐𝑜𝑙𝑜𝑟(𝐼𝑠𝑒𝑔) (12)
Shape descriptors are represented in Eq. (13):
𝐹𝑠ℎ𝑎𝑝𝑒 = [𝐴𝑟𝑒𝑎, 𝑃𝑒𝑟𝑖𝑚𝑒𝑡𝑒𝑟, 𝑀𝑎𝑗𝑜𝑟𝐴𝑥𝑖𝑠] (13)
These features are concatenated with deep
attention features to form the final representation, as
expressed in Eq. (14):
𝐹𝑡𝑜𝑡𝑎𝑙 = [𝐹𝑎𝑡𝑡, 𝐹𝑡𝑒𝑥𝑡𝑢𝑟𝑒, 𝐹𝑐𝑜𝑙𝑜𝑟, 𝐹𝑠ℎ𝑎𝑝𝑒] (14)
As shown in Eq. (14), combining multiple feature
types improves robustness and interpretability.
3.7 Feature selection
To eliminate redundant and irrelevant features, a
selection mechanism is applied, as defined in Eq.
(15):
𝐹𝑠𝑒𝑙 = 𝑆𝑒𝑙𝑒𝑐𝑡(𝐹𝑡𝑜𝑡𝑎𝑙) (15)
As indicated in Eq. (15), feature selection reduces
dimensionality and enhances generalization
performance. Fig. 3 shows the Illustration of
attention-based feature refinement, multi -level
feature extraction (texture, color, and shape), feature
fusion, and feature selection process.
Figure. 3 Efficient long-range attention mechanism and multi-level feature extraction, fusion, and selection framework
for cotton leaf disease classification
Received: September 4, 2026. Revised: October 3, 2026. 376
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
3.8 Adaptive fuzzy feature weighting
To address uncertainty in feature importance,
fuzzy logic is introduced. The membership value is
computed using Eq. (16):
𝜇𝑖 =
1
1+𝑒−𝐹𝑠𝑒𝑙,𝑖 (16)
The normalized weight is obtained using Eq. (17):
𝑤𝑖 = 𝜇𝑖
∑ 𝜇𝑗
𝑑
𝑗=1
(17)
Finally, the fuzzy -weighted feature
representation is calculated as in Eq. (18):
𝐹𝑓𝑢𝑧𝑧𝑦 = ∑ 𝑤𝑖
𝑑
𝑖=1
𝐹𝑠𝑒𝑙,𝑖 (18)
As described in Eqs. (16)–(18), the fuzzy module
dynamically adjusts feature contributions, improving
robustness under environmental variability.
3.9 Depthwise ShCNN classification
The fuzzy feature vector is processed using a
lightweight depthwise CNN. The convolution
operation is defined in Eq. (19):
𝑌 = 𝐶𝑜𝑛𝑣𝑝𝑜𝑖𝑛𝑡𝑤𝑖𝑠𝑒(𝐶𝑜𝑛𝑣𝑑𝑒𝑝𝑡ℎ𝑤𝑖𝑠𝑒(𝐹𝑓𝑢𝑧𝑧𝑦)) (19)
The activated feature map is obtained using Eq. (20):
𝑍 = 𝑅𝑒𝐿𝑈(𝐵𝑁(𝑌)) (20)
The final prediction is computed using the
Softmax classifier in Eq. (21):
𝑦̂ = 𝑆𝑜𝑓𝑡𝑚𝑎𝑥(𝑊𝑍 + 𝑏) (21)
As shown in Eqs. (19)–(21), depthwise separable
convolution reduces computational cost while
maintaining classification accuracy.
3.10 Loss function
The classification loss is defined using cross -
entropy, as shown in Eq. (22):
ℒ = − ∑ 𝑦𝑖
𝐶
𝑖=1
log (𝑦̂𝑖) (22)
To address class imbalance, a weighted loss is
used as in Eq. (23):
ℒ𝑤𝑒𝑖𝑔ℎ𝑡𝑒𝑑 = − ∑ 𝑤𝑖
𝐶
𝑖=1
𝑦𝑖log (𝑦̂𝑖) (23)
3.11 Bluefin trevally optimization algorithm
(BTOA) for hyperparameter tuning
To enhance model performance and ensure
optimal convergence, the proposed framework
incorporates the Bluefin Trevally Optimization
Algorithm (BTOA) for automatic hyperparameter
tuning. Let the set of hyperparameters be defined as
in Eq. (24):
Θ = {𝑙𝑟, 𝑏𝑎𝑡𝑐ℎ, 𝑑𝑟𝑜𝑝𝑜𝑢𝑡, 𝑤𝑑, 𝑓𝑖𝑙𝑡𝑒𝑟𝑠} (24)
where, 𝑙𝑟 = learning rate, 𝑏𝑎𝑡𝑐ℎrepresents the batch
size, 𝑑𝑟𝑜𝑝𝑜𝑢𝑡 corresponds to the dropout rate,
𝑤𝑑indicates weight decay, and 𝑓𝑖𝑙𝑡𝑒𝑟𝑠specifies the
number of convolutional filters. The objective of
BTOA is to determine the optimal parameter
configuration that minimizes the classification loss,
as expressed in Eq. (25):
Θ∗ = arg min
Θ
ℒ(Θ) (25)
Initially, a population of candidate solutions is
generated, where each solution represents a
hyperparameter configuration, as defined in Eq. (26):
𝑋𝑖(0) = [𝜃1, 𝜃2, . . . , 𝜃𝑑] (26)
The position of every candidate solution is
updated iteratively according to both exploration and
exploitation-based strategies during optimization
process. Each candidate solution is evaluated using
the weighted loss function defined earlier in Eqs. (22)
and (23), and the fitness is computed as in Eq. (27):
𝐹𝑖𝑡𝑛𝑒𝑠𝑠(𝑋𝑖) = ℒ𝑤𝑒𝑖𝑔ℎ𝑡𝑒𝑑 (27)
The optimal solution based on minimum fitness
value is selected according to Eq. (28):
𝑋𝑏𝑒𝑠𝑡 = arg min 𝑖 𝐹𝑖𝑡𝑛𝑒𝑠𝑠(𝑋𝑖) (28)
The BTOA incorporates automated
hyperparameter tuning, which leads to faster
convergence and better generalization. BTOA
explores the search space more systematically
Received: September 4, 2026. Revised: October 3, 2026. 377
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
compared to manual tuning approaches and filters
optimal configurations in a data-driven fashion.
BTOA hyperparameter optimization was
performed exclusively using the training and
validation partitions. Candidate models were fitted
using the training partition, while the validation
Macro-F1 was used as the optimization criterion for
candidate ranking a nd hyperparameter selection.
Model checkpoints were selected according to
validation performance without accessing the
independent test labels or test predictions. After the
hyperparameter configuration had been fixed, the
selected configuration was evalua ted once on the
held-out test partition for final performance reporting.
The test partition did not contribute to BTOA fitness
computation, checkpoint selection, early stopping, or
hyperparameter selection.
After completion of BTOA, the hyperparameter
configuration yielding the highest validation Macro -
F1 was fixed. A final model was subsequently trained
on the original training partition using this
configuration, while the validation partition was
retained for early stopping and checkpoint selection.
Only after completion of this final model -selection
procedure was the independent test partition
evaluated once for final performance reporting.
3.12 Implementation details and reproducibility
To ensure complete reproducibility and address
reviewer concerns regarding experimental
transparency, all experiments were conducted under
a common and well -defined experimental
configuration. The proposed EfficientNet –Fuzzy
ShCNN framework was implemented using the
PyTorch deep learning framework and executed on a
workstation equipped with an Intel Xeon processor,
64 GB RAM, and an NVIDIA RTX 3090 (24 GB) or
Tesla T4 GPU with CUDA acceleration. A common
set of hyperparameters was adopted for all
experiments, including 120 training epochs, a batch
size of 64, a learning rate of 1 × 10−4, the Adam
optimizer, weight decay of 1 × 10−5, and a dropout
rate of 0.3. The parameter update rule of the Adam
optimizer is given in Eq. (30):
𝜃𝑡+1 = 𝜃𝑡 − 𝛼
𝑚𝑡
√𝑣𝑡+𝜖 (30)
where 𝛼 denotes the learning rate, and 𝑚𝑡 and
𝑣𝑡 represent the first - and second -order moment
estimates, respectively. Early stopping with a
patience of 15 epochs was employed to prevent
overfitting, and the ShCNN classifier was regularized
using dropout to improve generalization performance.
Dataset composition: The publicly available Cotton
Leaf Disease Dataset (Version 1) was obtained from
Kaggle. The source dataset comprises four classes:
Healthy, Bacterial Blight, Leaf Curl Virus, and
Fusarium Wilt. Following dataset screening, the
experimental subset used in this study contained
1,400 images. The detailed class -wise dataset
composition and partitioning are presented in Table
2.
The source Cotton Leaf Disease dataset contained
1,711 images across the four considered classes.
Following dataset screening and experimental-subset
construction, 1,400 images were retained for the
experiments, with 350 images assigned to each class.
The difference between the source -dataset size and
the retained experimental subset is not interpreted as
the number of duplicate images, because separate
counts of exact -duplicate and near -duplicate
exclusions were not retained. The retained images
were scree ned for duplicate and near -duplicate
instances and assigned unique source -image
identifiers before data partitioning. Prior to any data
augmentation, the 1,400 -image experimental subset
was stratified into mutually exclusive training (70%;
980 images), val idation (15%; 210 images), and
testing (15%; 210 images) subsets using a fixed
dataset-partition seed of 42. A split manifest
containing the source -image identifier, class label,
and assigned subset was maintained to ensure that
each retained original imag e occurred in only one
subset. Data augmentation was subsequently applied
exclusively to the training subset, whereas the
validation and test subsets contained only original,
non-augmented images.
The identical dataset partition was used for all
baseline models and the proposed EfficientNet –
Fuzzy ShCNN framework. To quantify variability
Table 2. Dataset Composition and Leakage-Free Data Partition
Class Source Dataset Images Images Retained for Experimental Subset Train Validation Test
Healthy 426 350 245 52 53
Bacterial Blight 448 350 245 53 52
Leaf Curl Virus 418 350 245 52 53
Fusarium Wilt 419 350 245 53 52
Total 1,711 1,400 980 210 210
Received: September 4, 2026. Revised: October 3, 2026. 378
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
arising from stochastic model training, each
architecture was independently trained five times
using training seeds 42, 52, 62, 72, and 82. For each
run, the corresponding training seed -controlled
weight initialization, mini -batch shuffling, dropout,
and stochastic training-time augmentation. The same
five training seeds were used for every architec ture,
enabling seed-wise paired statistical comparisons. All
comparative models were evaluated using the same
image resolution, preprocessing protocol, training -
only augmentation policy, dataset partitions, and
evaluation metrics. The reported results correspond
to the mean and standard deviation obtained across
the five independent training runs. The complete
experimental configuration is summarized in Table
2A.
4. Experimental results
A novel multi -level classifier for cotton leaf
diseases based on the EfficientNet –Fuzzy ShCNN
pipeline was proposed using a Cotton Disease
Dataset (Kaggle) as input. The proposed model is
realized using Python libraries OpenCV for image
pre-processing, NumPy for numerical computations,
and PyTorch for deep learning based feature
extraction and classification. The computations have
been performed on a high-performance type machine
with Intel Xeon CPU, 64 GB RAM and NVIDIA
RTX 3090 or Tesla T4 GPU to enhance the
computational speed. The dataset contains
approximately 1,400 original images distributed
across four classes: Healthy, Bacterial Blight, Leaf
Table 2A. Experimental Configuration for
Reproducibility
Parameter Value
Framework PyTorch 2.x
Python 3.x
GPU NVIDIA RTX 3090 (24 GB) /
Tesla T4
CPU Intel Xeon
RAM 64 GB
Image Resolution 224 × 224
Dataset Split Stratified 70% / 15% / 15%
Dataset Partition Seed 42
Training Seeds 42, 52, 62, 72, 82
Optimizer Adam
Learning Rate 1 × 10⁻⁴
Weight Decay 1 × 10⁻⁵
Batch Size 64
Epochs 120
Early Stopping Patience = 15 epochs
Independent Training
Runs
5
Evaluation Mean ± standard deviation over
five runs
Curl Virus, and Fusarium Wilt. The original samples
were divided into training, validation, and test subsets
before augmentation.
4.1 Performance analysis
To ensure a fair and unbiased comparison, all
baseline models and the proposed EfficientNet –
Fuzzy ShCNN framework were trained and evaluated
under an identical experimental protocol. The same
stratified dataset partition (70:15:15), preprocessing
operations, data augmentation strategy, optimizer
configuration, learning rate, batch size, weight decay,
dropout rate, early -stopping criterion, inpu t image
resolution, and evaluation metrics were applied
consistently across all models, as summarized in
Table 3. Each architecture was independently trained
and evaluated over five experimental runs using the
same set of five training seeds.
Table 4 presents the comparative performance of
the proposed EfficientNet–Fuzzy ShCNN framework
against conventional and advanced deep learning
architectures under the same experimental protocol.
The baseline CNN achieves an accuracy of 92.14%,
reflecting its comparatively limited feature -
representation capability. ResNet50 improves the
accuracy to 94.82%, benefiting from residual
connections that facilitate gradient propagation and
support deeper feature extraction. EfficientNet -B0
further increases the ac curacy to 96.21% through its
compound scaling strategy, which balances network
depth, width, and input resolution.
Integrating the attention mechanism with
EfficientNet-B0 improves the accuracy to 97.42%,
indicating that modelling long -range feature
dependencies and emphasizing disease -relevant
regions can improve classification performance. The
proposed EfficientNet –Fuzzy ShCNN framework
achieves the highest accuracy of 98.72%, together
with precision, specificity, recall, and F1 -score
values exceeding 98%.
The class -wise performance of the proposed
EfficientNet–Fuzzy ShCNN framework is presented
in Table 5. The Leaf Curl Virus class achieves the
highest F1-score of 99.07%, followed closely by the
Healthy class with an F1 -score of 99.05%. The
Bacterial Blight and Fusarium Wilt classes each
achieve an F1-score of 98.08%. These results indicate
that the proposed framework maintains consistently
high classification performance across all four
categories. The confusion matrix shown in Fig. 4 was
generated from the same representative test -run
predictions used to compute the class -wise metrics
reported in Table 5, thereby ensuring numerical
consistency between the two evaluations.
Received: September 4, 2026. Revised: October 3, 2026. 379
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
Table 3. Training Configuration Used for All Comparative Models
Configuration CNN ResNet50 EfficientNet-
B0
EfficientNet +
Attention
Proposed
Training Epochs 120 120 120 120 120
Optimizer Adam Adam Adam Adam Adam
Learning Rate 1 × 10⁻⁴ 1 × 10⁻⁴ 1 × 10⁻⁴ 1 × 10⁻⁴ 1 × 10⁻⁴
Batch Size 64 64 64 64 64
Weight Decay 1 × 10⁻⁵ 1 × 10⁻⁵ 1 × 10⁻⁵ 1 × 10⁻⁵ 1 × 10⁻⁵
Dropout Rate 0.30 0.30 0.30 0.30 0.30
Early Stopping Patience = 15 Patience =
15
Patience = 15 Patience = 15 Patience =
15
Input Image Size 224 × 224 224 × 224 224 × 224 224 × 224 224 × 224
Dataset Split Same 70:15:15 Same Same Same Same
Data
Augmentation
Same Same Same Same Same
Number of Runs 5 5 5 5 5
Evaluation
Metrics
Accuracy, Precision,
Recall, F1, AUC
Same Same Same Same
Table 4. Performance Analysis with Traditional Models
Performance Models Accuracy (%) Precision (%) Specificity (%) F1-score (%) Recall (%)
CNN 92.14 ± 0.28 91.80 ± 0.26 92.05 ± 0.27 91.72 ± 0.25 91.65 ± 0.24
ResNet50 94.82 ± 0.26 94.30 ± 0.25 94.55 ± 0.24 94.20 ± 0.23 94.10 ± 0.24
EfficientNet 96.21 ± 0.24 95.90 ± 0.23 96.10 ± 0.22 95.85 ± 0.23 95.80 ± 0.22
EfficientNet + Attention 97.42 ± 0.22 97.10 ± 0.21 97.30 ± 0.22 97.05 ± 0.21 97.00 ± 0.21
Proposed EfficientNet–Fuzzy
ShCNN 98.72 ± 0.21 98.50 ± 0.20 98.65 ± 0.21 98.55 ± 0.20 98.60 ± 0.21
Table 5. Analysis of Class-wise Performance
Class Precision (%) Recall (%) F1-score (%)
Healthy 100.00 98.11 99.05
Bacterial
Blight 98.08 98.08 98.08
Leaf Curl
Virus 98.15 100.00 99.07
Fusarium
Wilt 98.08 98.08 98.08
Macro
Average 98.58 98.57 98.57
Table 5 presents the class -wise performance of
the proposed EfficientNet–Fuzzy ShCNN framework
for a representative test run, while Fig. 4 shows the
corresponding confusion matrix generated from the
same set of predictions. The representative run was
selected as the run whose overall F1 -score was
closest to the five-run mean reported in Table 4. The
independent test subset contained 210 original, non -
augmented images, with no augmented samples
included during evaluation. The model correctly
Figure. 4 Confusion matrix of the proposed EfficientNet–
Fuzzy ShCNN model for a representative run on the
independent non-augmented test set
classified 207 of the 210 test images, corresponding
to an accuracy of 98.57%. The precision, recall, and
F1-score values reported in Table 5 were computed
Received: September 4, 2026. Revised: October 3, 2026. 380
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
Figure. 5 Error rates comparison of various models on
cotton leaf disease classification
directly from the row and column totals of the
confusion matrix shown in Fig. 4. In contrast, Table
4 reports the mean ± standard deviation obtained
across five independent training runs.
Table 7 presents a controlled head -to-head
comparison with Fine -tuned VGG16, Swin
Transformer-T, and EcoNet, which were
reimplemented using the same dataset partition,
training-only augmentation policy, input resolution,
and evaluation protocol as the proposed framework.
Among the comparative methods, the Swin
Transformer achieved strong classification
performance with an accuracy of 98.31% and macro-
F1 of 98.18%, while EcoNet obtained 98.14%
accuracy and 98.02% macro -F1. The proposed
EfficientNet–Fuzzy ShCNN framework achieved the
highest accuracy of 98.72%, macro -F1 of 98.55%,
and balanced accuracy of 98.63%. It also obtained the
lowest ECE of 1.28%, indicating improved
calibration. Moreover, the proposed model required
only 4.87 M parameters and 0.46 GFLOPs, providing
a favorable trade -off between classification
performance and computational complexity
compared with the evaluated recent architectures.
All comparative architectures were evaluated
under a common controlled experimental protocol
using the same 1,400 -image balanced dataset, fixed
70/15/15 train-validation-test partition, training-only
augmentation, optimization settings, early -stopping
rule, checkpoint-selection criterion, and five matched
random seeds (42, 52, 62, 72, and 82). Model
selection was performed exclusively using validation
Macro-F1, and the test set was not used for
Table 6. Literature-Level Performance Summary of Existing Cotton Disease Classification Methods
Study Model / Approach Reported Performance
Hu et al. [5] ACURS-YOLO / YOLOv11-based mAP@0.5 = 94.6%
Prasad et al. [6] Multi-Stream Attention-Guided Hybrid CNN Accuracy = 97.32%; F1 ≈ 98%
Amin et al. [7] Explainable VGG-16-based network Accuracy = 99.99%
Rai and Pahuja [11] Improved Deep CNN Accuracy = 97.98%
Arathi and Dulhare [12] DenseNet-121 Transfer Learning Accuracy = 91%
Kaur et al. [13] Fine-tuned VGG16 Accuracy = 95.5%
Pandey et al. [14] Hybrid Deep Learning Model Accuracy = 98.9%
Rajasekar et al. [15] ResNet-50 + Xception Validation Accuracy = 98%
Proposed EfficientNet –Fuzzy
ShCNN
Hybrid DL + Fuzzy ShCNN Accuracy = 98.72%
Table 7. Head-to-Head Comparison with Recent Cotton Leaf Disease Classification Models under an Identical
Experimental Protocol
Method Year Accuracy (%) Macro-F1 (%) Balanced Accuracy (%) Mean ECE (%) across 5 runs Parameters (M) FLOPs (G) Latency (ms)
Fine-tuned
VGG16 2024 96.84 96.71 96.78 2.86 14.72* 15.35* 6.94*
Swin Transformer-
T 2025 98.31 98.18 98.24 1.61 28.29 4.50 7.82
EcoNet 2025 98.14 98.02 98.08 1.74 6.41 0.71 4.68
EfficientNet-B0 – 96.21 95.85 96.03 2.31 5.30 0.39 3.60
EfficientNet-B0 +
Attention – 97.42 97.05 97.22 1.92 6.12 0.52 4.06
Proposed
EfficientNet–
Fuzzy ShCNN
– 98.72 98.55 98.63 1.28 4.87 0.46 3.79
Received: September 4, 2026. Revised: October 3, 2026. 381
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
Table 7A. Reimplementation Configuration of Models Used in the Controlled Head -to-Head Comparison
Configuration Fine-tuned
VGG16
Swin
Transformer-T
EcoNet EfficientNet-B0 EfficientNet-B0
- Attention
Implementation
framework
PyTorch PyTorch PyTorch PyTorch PyTorch
Python version 3.10.13 3.10.13 3.10.13 3.10.13 3.10.13
PyTorch version 2.2.2 2.2.2 2.2.2 2.2.2 2.2.2
torchvision
version
0.17.2 0.17.2 0.17.2 0.17.2 0.17.2
CUDA version 12.1 12.1 12.1 12.1 12.1
cuDNN version 8.9.2 8.9.2 8.9.2 8.9.2 8.9.2
Implementation
source
torchvision
VGG16
torchvision
Swin-T
Custom
PyTorch EcoNet
reconstruction
torchvision
EfficientNet-B0
torchvision
EfficientNet-B0 - custom long -
range attention
Architecture
variant
VGG16 Swin-T EcoNet EfficientNet-B0 EfficientNet-B0 - long -range
attention
Input
configuration
224 × 224 × 3
RGB
224 × 224 × 3
RGB
224 × 224 × 3
RGB
224 × 224 × 3
RGB
224 × 224 × 3
RGB
Output classes 4 4 4 4 4
Classifier
adaptation
Final classifier
adapted to 4
classes
Final
classification
head adapted to
4 classes
Terminal
classifier
adapted to 4
classes
Final
classification
head adapted to
4 classes
Attention-
enhanced
representation →
4-class classifier
Fine-tuning
configuration
End-to-end fine -
tuning
End-to-end fine -
tuning
End-to-end
training
End-to-end fine -
tuning
End-to-end fine -
tuning of
backbone,
attention and
classifier
Parameters (M) 14.72 28.29 6.41 5.30 6.12
FLOPs (G) 15.35 4.50 0.71 0.39 0.52
Optimizer Adam Adam Adam Adam Adam
Initial learning
rate
1 × 10⁻⁴ 1 × 10⁻⁴ 1 × 10⁻⁴ 1 × 10⁻⁴ 1 × 10⁻⁴
Weight decay 1 × 10⁻⁵ 1 × 10⁻⁵ 1 × 10⁻⁵ 1 × 10⁻⁵ 1 × 10⁻⁵
Batch size 64 64 64 64 64
Maximum
epochs
120 120 120 120 120
Early-stopping
patience
15 epochs 15 epochs 15 epochs 15 epochs 15 epochs
Checkpoint-
selection
criterion
Highest validation
Macro-F1
Highest
validation
Macro-F1
Highest
validation
Macro-F1
Highest
validation
Macro-F1
Highest
validation Macro-
F1
Training
augmentation
Rotation ±20°;
horizontal flip
p=0.5; brightness
±15%; scaling
0.9–1.1×
Same Same Same Same
Dataset size 1,400 images 1,400 images 1,400 images 1,400 images 1,400 images
Number of
classes
4 4 4 4 4
Images per class 350 350 350 350 350
Dataset split 70%/15%/15% 70%/15%/15% 70%/15%/15% 70%/15%/15% 70%/15%/15%
Training images 980 980 980 980 980
Validation
images
210 210 210 210 210
Test images 210 210 210 210 210
Partition seed 42 42 42 42 42
Received: September 4, 2026. Revised: October 3, 2026. 382
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
Training seeds 42, 52, 62, 72, 82 42, 52, 62, 72, 82 42, 52, 62, 72,
82
42, 52, 62, 72, 82 42, 52, 62, 72, 82
Independent
runs
5 5 5 5 5
Model-selection
data
Validation set only Validation set
only
Validation set
only
Validation set
only
Validation set
only
Test-set role in
selection
None None None None None
Accuracy (%) 96.84 98.31 98.14 96.21 97.42
Macro-F1 (%) 96.71 98.18 98.02 95.85 97.05
Balanced
Accuracy (%)
96.78 98.24 98.08 96.03 97.22
Mean ECE (%) 2.86 1.61 1.74 2.31 1.92
Table 7B. Seed-Wise Macro-F1 Performance of the Controlled Reimplemented Models under the Common Experimental
Protocol
Model Seed 42 Seed 52 Seed 62 Seed 72 Seed 82 Mean ± SD
Fine-tuned VGG16 96.82 96.55 96.91 96.64 96.63 96.71 ± 0.15
Swin Transformer-T 98.31 98.07 98.28 98.16 98.08 98.18 ± 0.11
EcoNet 98.14 97.91 98.17 97.96 97.92 98.02 ± 0.12
EfficientNet-B0 95.93 96.00 95.54 95.68 96.09 95.85 ± 0.23
EfficientNet-B0 + Attention 96.83 97.32 96.99 96.89 97.22 97.05 ± 0.21
Proposed EfficientNet–Fuzzy ShCNN 98.70 98.78 98.54 98.46 98.28 98.55 ± 0.20
hyperparameter tuning or checkpoint selection.
Model-specific changes were restricted to the
adaptations required for four-class classification and,
for EfficientNet-B0 + Attention, incorporation of the
long-range attention module. Individual seed -level
results are reported in Table 7B.
4.2 Comparison with recent techniques
Table 6 summarizes performance values reported
in representative cotton leaf disease classification
studies and is included only for literature -level
contextualization because these studies employ
different datasets, sample sizes, preprocessing
strategies, and evaluation protocols. Table 7 presents
a controlled head -to-head comparison with Fine -
tuned VGG16, Swin Transformer -T, and EcoNet,
which were reimplemented using the same dataset
partition, training -only augmentation policy, input
resolution, and eval uation protocol as the proposed
framework.
4.3 Ablation study
To isolate the individual contribution of each
component, a controlled leave -one-component-out
ablation study was performed. Starting from the
complete EfficientNet –Fuzzy ShCNN framework,
Gaussian filtering, FLICM segmentation,
handcrafted descriptors, feature selection, long-range
attention, fuzzy feature weighting, and the ShCNN
classifier were independently removed or replaced
while all remaining training conditions were held
constant. The contribution of this formulation is
evaluated through controlled component-wise
ablation rather than inferred solely from comparisons
with previously published accuracy values.
Specifically, removal of long -range attention,
handcrafted descriptors, feature selection, fuzzy
weighting, ShCNN, preprocessing stages, and BTOA
optimization is evaluated while the remaining
experimental conditions are held fixed. The BTOA
component is additionally compared with alternative
hyperparameter-selection strategies under an equal
candidate-evaluation budget. Consequently, the
empirical contribution of the proposed framework is
attributed to the coordinated feature -interaction and
classification design under the evaluated benchmark
protocol rather than to the independent novelty of its
constituent techniques. The results are summarized in
Table 8A as mean ± standard deviation across five
independent training seeds. In addition, the
contribution of BTOA was evaluated against manual
configuration, random search, and Bayesian
optimization under an equal hyperparameter -search
budget, as reported in Table 8B.
During BTOA optimization, each candidate
hyperparameter configuration was trained
exclusively on the training partition and evaluated on
the validation partition. Validation Macro -F1 served
as the BTOA fitness objective. Early stopping was
performed within each candidate training process
using validation performance, and the checkpoint
achieving the highest validation Macro -F1 was
Received: September 4, 2026. Revised: October 3, 2026. 383
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
Table 8A. Component-Wise Ablation Analysis of the Proposed EfficientNet–Fuzzy ShCNN Framework
Configuration Accuracy
(%)
Macro-F1
(%)
Δ Accuracy
(%)
Parameters
(M)
End-to-End
Latency (ms)
Full Proposed Model 98.72 ± 0.21 98.55 ± 0.20 – 4.87 14.64
Without Gaussian Filtering 98.31 ± 0.22 98.14 ± 0.24 −0.41 4.87 14.12
Without FLICM Segmentation 97.86 ± 0.25 97.68 ± 0.27 −0.86 4.87 9.98
Without Handcrafted Features 97.94 ± 0.24 97.76 ± 0.26 −0.78 4.61 12.49
Without Feature Selection 98.09 ± 0.23 97.91 ± 0.25 −0.63 4.92 14.08
Without Long-Range Attention 97.43 ± 0.27 97.05 ± 0.21 −1.29 4.64 14.11
Without Fuzzy Weighting 97.96 ± 0.26 97.69 ± 0.28 −0.76 4.87 14.37
Replace ShCNN with Standard
CNN Head
97.71 ± 0.28 97.46 ± 0.29 −1.01 5.36 15.28*
Without BTOA Optimization 97.84 ± 0.25 97.61 ± 0.27 −0.88 4.87 14.64
Table 8B. Comparison of Hyperparameter Optimization Strategies under an Equal Search Budget
Optimization
Method
Validation
Macro-F1 (%)
Test Accuracy
(%)
Test Macro -
F1 (%)
Candidate
Evaluations
Optimization
Time (h)
Manual Baseline 97.49 97.84 97.61 600 7.4
Random Search 97.76 98.05 97.87 600 8.1
Bayesian
Optimization
98.03 98.39 98.20 600 9.3
BTOA 98.37 98.72 98.55 600 8.7
retained for candidate evaluation. Consequently,
neither test labels nor test predictions contributed to
candidate fitness computation, early stopping,
checkpoint selection, candidate ranking, or
termination of the BTOA search. A total of 600
candidate con figurations were evaluated under the
predefined search budget.
The specific hyperparameter settings utilized in
the proposed framework are summarized in Table 9.
To ensure stable convergence during training, the
Adam optimizer was used with a learning rate of
1 × 10−4 and weight decay of 1 × 10−5 . Dropout
regularization and early stopping were used to reduce
overfitting. Hyperparameter configurations were
selected exclusively according to validation Macro -
F1.
For reproducibility, Table 9 provides the
complete dimensional configuration of the proposed
framework. The handcrafted branch initially
produces an 87 -dimensional representation
comprising 40 LGBP, 32 WLD, 12 color, and three
shape descriptors. Mutual -information ranking
retains the 64 most informative features, which are
standardized using statistics estimated exclusively
from the training subset and subsequently projected
to 128 dimensions. In parallel, the 1280-dimensional
EfficientNet-B0 representation is projected to the
same 128 -dimensional latent space. Sigmoid -based
fuzzy memberships followed by softmax
normalization are used to adaptively weight the two
representations before element -wise fusion. The
resulting 128 -dimensional fused representation i s
processed by the depthwise -separable ShCNN
classifier through 64 - and 32 -dimensional
intermediate representations before four -class
prediction. To ensure transparency and
reproducibility of the statistical evaluation, the seed-
wise F1 -scores obtained fro m the five independent
training runs are presented in Table 10. The same
training seeds (42, 52, 62, 72, and 82) were used for
all evaluated architectures, enabling direct pairing of
corresponding runs. The reported mean and standard
deviation were calcula ted across the five runs for
each model. All four paired comparisons were highly
statistically significant ( 𝑝 < 0.001), indicating that
the proposed EfficientNet–Fuzzy ShCNN framework
achieved significantly higher F1 -scores than the
evaluated baseline architectures.
All statistics reported in Table 10 were calculated
from the same unrounded five run -level F1 -score
observations.
Statistical significance was assessed at 𝛼 = 0.05.
All four paired comparisons were highly statistically
significant (𝑝 < 0.001), indicating that the proposed
EfficientNet–Fuzzy ShCNN framework achieved
significantly higher F1 -scores than the evaluated
baseline architectures. Table 10A presents the paired
statistical comparisons based on the five independent
seed-wise training runs. All statistics were computed
from the same unrounded run -level F1 -score
observations, with corresponding runs paired
according to their training seeds. The reported mean
paired differences, standard deviations of the paired
differences, t -statistics, p -values, and 95%
confidence intervals were therefore derived from the
Received: September 4, 2026. Revised: October 3, 2026. 384
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
Table 9. Comprehensive Hyperparameter Configuration for Training the Proposed EfficientNet-Fuzzy ShCNN
Framework
Component Parameter Value
Input / Pre-processing Input image size 224 × 224 × 3
Normalization Min–max scaling to [0, 1]
Augmentation Rotation (±20°), horizontal flip (𝑝 = 0.5), brightness
(±15%), scaling (0.9–1.1×)
Gaussian Filtering Kernel size 5 × 5
Standard deviation (𝜎) 1.0
FLICM Segmentation Number of clusters (𝐶) 2
Fuzziness coefficient (𝑚) 2.0
Spatial weight (𝜆) 0.5
Neighborhood 3 × 3, 8-connected
Cluster-selection rule Cluster with greater leaf-region occupancy after
foreground-mask extraction
EfficientNet Backbone Backbone EfficientNet-B0, ImageNet pretrained
Input channels 3
Original classification
head
Removed
Backbone feature
dimension
1280-D
Long-Range Attention Attention heads 2
Projected/hidden
dimension
128-D
LGBP Features Gabor scales ×
orientations
5 × 8 = 40
Output dimension 40-D
WLD Features Histogram descriptor
dimension
32-D
Color Features RGB statistics 6-D (mean + SD for 3 channels)
HSV statistics 6-D (mean + SD for 3 channels)
Total RGB + HSV
dimension
12-D
Shape Features Descriptors Area, perimeter, major-axis length
Output dimension 3-D
Handcrafted Features Combined feature
dimension
87-D
Feature Selection Method Mutual-information ranking
Number of retained
features (𝐾)
64
Fitting strategy Training subset only
Feature Normalization Method Z-score standardization using training-set mean and SD
Deep-Feature Projection Input → output 1280-D → 128-D
Handcrafted-Feature
Projection
Input → output 64-D → 128-D
Fuzzy Module Membership function Sigmoid
Sigmoid slope (𝑎) 1.0
Sigmoid center (𝑐) 0.0
Weight normalization Softmax normalization
Feature Fusion Fusion strategy Element-wise fuzzy weighted fusion
Deep representation 128-D
Handcrafted
representation
128-D
Fused output dimension 128-D
ShCNN Classifier Input tensor 128 × 1 × 1
Convolution type Depthwise separable
First block 128 → 64 channels
Second block 64 → 32 channels
Received: September 4, 2026. Revised: October 3, 2026. 385
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
Intermediate
representation
64-D → 32-D
Activation ReLU
Dropout 0.30
Output Layer Input dimension 32-D
Number of classes 4
Activation Softmax
Training Optimizer Adam
Learning rate 1 × 10−4
Weight decay 1 × 10−5
Batch size 64
Maximum epochs 120
Early-stopping patience 15 epochs
BTOA Population size 20
Number of iterations 30
Candidate evaluations 600
Optimization criterion Validation Macro-F1
Hyperparameter-selection
data
Training + validation only
Experimental Protocol Data split 70% / 15% / 15%
Partition seed 42
Independent training seeds 42, 52, 62, 72, 82
Table 10. Seed-Wise Run-Level F1-Scores of the Proposed Framework and Baseline Models
Seed CNN ResNet-50 EfficientNet-B0 EfficientNet-B0 + Attention Proposed
42 91.5946 94.4278 95.9331 96.8328 98.6981
52 91.8498 93.9076 96.0006 97.3183 98.7816
62 91.4287 94.4404 95.5402 96.9936 98.5392
72 92.0773 94.1338 95.6842 96.8882 98.4556
82 91.6497 94.0904 96.0919 97.2170 98.2754
Mean 91.72 94.20 95.85 97.05 98.55
Table 10A. Paired Statistical Comparison Based on Five Independent Runs
Model Comparison Mean F1 Difference (%) SD of Paired Differences (%) t-statistic df p-value 95% CI
Proposed vs CNN +6.83 0.320 47.70 4 0.0000012 [6.43, 7.23]
Proposed vs ResNet50 +4.35 0.305 31.91 4 0.0000057 [3.97, 4.73]
Proposed vs EfficientNet-
B0 +2.70 0.305 19.81 4 0.000038 [2.32, 3.08]
Proposed vs EfficientNet +
Attention +1.50 0.290 11.57 4 0.000319 [1.14, 1.86]
same underlying observations, ensuring consistency
between the significance tests and confidence
intervals.
4.4 Computational complexity analysis
To evaluate the computational efficiency of the
proposed framework, its complexity was compared
with representative deep learning models in terms of
trainable parameters, floating -point operations
(FLOPs), inference speed, average inference latency,
and GP U memory consumption. All models were
evaluated under the same hardware environment
using an NVIDIA RTX 3090 GPU with 24 GB
memory and an identical input resolution of 224 ×
224pixels to ensure a consistent comparison.
As shown in Table 11, the proposed EfficientNet–
Fuzzy ShCNN framework requires 4.87 M trainable
parameters and 0.46 GFLOPs, with an average
inference latency of 3.79 ms/image, corresponding to
approximately 264 FPS. The lower parameter count
relative to the standard EfficientNet-B0 model should
not be interpreted as adding the proposed modules to
the complete 5.30 M -parameter EfficientNet -B0
architecture.
Received: September 4, 2026. Revised: October 3, 2026. 386
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
Table 11. Computational Complexity Comparison
Model Parameters (M) FLOPs (G) FPS Inference Time (ms/image) GPU Memory (GB)
CNN 11.20 3.54 226 4.42 1.21
ResNet50 25.60 4.12 162 6.18 2.43
EfficientNet-B0 5.30 0.39 278 3.60 1.08
EfficientNet-B0 + Attention 6.12 0.52 246 4.06 1.32
Proposed EfficientNet–Fuzzy
ShCNN 4.87 0.46 264 3.79 1.18
The use of depthwise separable convolution in the
ShCNN classifier further limits the number of
additional trainable parameters while maintaining the
discriminative capability of the extracted features.
Consequently, the resulting architecture provides a
favorable balance between classification
performance and computational complexity under the
evaluated GPU environment. A component -wise
parameter analysis is provided in Table 11A to
explicitly show the contribution of each trainable
module to the overall parameter count.
The FPS values reported in Table 11 were derived
from the measured average inference latency
according to 𝐹𝑃𝑆 = 1000/𝑇ms , where
𝑇msrepresents the average inference time per image
in milliseconds. Thus, the reported FPS values
represent latency -derived single -image throughput
rather than an independently measured batched -
throughput experiment.
4.4.1. Component-wise parameter analysis
To clarify the parameter efficiency of the
proposed architecture, a component -wise parameter
analysis was performed. The standard EfficientNet -
B0 parameter count includes its original ImageNet
classification head. In the proposed framework, this
classification head is removed and EfficientNet-B0 is
retained only as the convolutional feature -extraction
backbone. The extracted representation is
subsequently processed by the proposed long -range
attention, feature -fusion, fuzzy -weighting, and
lightweight depthwise ShCNN modules. Therefore,
the parameter count of the proposed framework
should not be interpreted as the complete
EfficientNet-B0 model plus additional modules.
Instead, it represents the truncated EfficientNet
feature backbone together with the newly introduced
trainable modules. The component -wise trainable
parameter counts obtained directly from the
implemented network are summarized in Table 11A.
Table 11A provides a component-wise analysis of
the trainable parameters of the proposed architecture.
After removal of the original ImageNet classification
head, the EfficientNet -B0 feature -extraction
backbone accounts for approximately 4.008 M
trainable parameters, representing 82.30% of the
complete architecture. The Efficient Long -Range
Attention module contributes approximately 0.230 M
parameters, while the feature -projection and multi -
level fusion layers account for approximately 0.350
M parameters. The adaptive fuzzy feature-weighting
operation does not introduce additional trainable
parameters because feature memberships and
normalized weights are computed directly from the
extracted representation. The lightweight depthwise
ShCNN classifier contributes approximately 0.282 M
parameters. Consequently, the complete proposed
framework contains approximately 4.870 M trainable
parameters, consistent with the overall parameter
count reported in Table 11.
4.4.2. End-to-end computational efficiency
To provide a complete assessment of
computational efficiency, the proposed framework
was profiled from raw image input to final disease
prediction. In contrast to model -only latency, this
analysis includes preprocessing, FLICM
segmentation, handcrafted feature extraction, feature
transformation, device transfer, neural -network
Table 11A. Component-Wise Trainable Parameter Distribution of the Proposed EfficientNet–Fuzzy ShCNN Framework
Component Trainable Parameters (M) Contribution (%)
EfficientNet-B0 feature backbone (classification head removed) 4.008 82.30
Efficient Long-Range Attention Module 0.230 4.72
Feature Projection and Multi-Level Fusion Layers 0.350 7.19
Adaptive Fuzzy Feature Weighting Module 0.000 0.00
Depthwise ShCNN Classifier 0.282 5.79
Total Proposed Framework 4.870 100.00
Received: September 4, 2026. Revised: October 3, 2026. 387
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
processing, and output generation. The module -wise
computational costs are reported in Table 12.
As shown in Table 12, the neural -network
components require 3.79 ms per image, whereas the
complete raw -image-to-prediction pipeline requires
14.64 ms. Among the preprocessing operations,
FLICM segmentation represents the largest
computational overhead. This analysis distinguishes
the previously reported neural inference latency from
the complete pipeline latency and provides a more
comprehensive assessment of computational
efficiency. Since BTOA operates during model
development rather than inference, its computational
overhead was separately quantified. Table 13
summarizes the BTOA search budget and training -
time computational requirements.
Table 13 shows that the BTOA optimization stage
required 600 candidate evaluations and
approximately 8.70 h under the evaluated
computational environment. This optimization is
performed offline and does not contribute to the
latency of the final trained model. As shown in Table
14, a statistically significant difference was observed
among the evaluated models, 𝐹(4, 20) = 688.45 ,
𝑝 < 0.0001.
4.5 Paired statistical analysis across training seeds
Because all five architectures were evaluated
using the same five training seeds, the resulting
observations constitute a paired repeated -measures
design rather than independent groups. Therefore, the
conventional one-way ANOVA was replaced by the
non-parametric Friedman test, which accounts for the
matched seed -wise evaluations. The analysis was
performed using the run -level Macro -F1 scores
reported in Table 10.
The architectures exhibited the same performance
ordering for all five seeds, resulting in mean ranks of
1.00, 2.00, 3.00, 4.00, and 5.00 for CNN, ResNet-50,
EfficientNet-B0, EfficientNet -B0 + Attention, and
the proposed framework, respectively.
Thus, the Friedman test indicates an overall
difference in the seed -wise Macro -F1 distributions
Table 12. Module-Wise and End-to-End Computational Latency of the Proposed Framework
Pipeline Stage Device Mean Latency (ms/image)
Image loading, resizing and normalization CPU 1.18
Gaussian filtering CPU 0.52
FLICM segmentation CPU 4.66
Handcrafted feature extraction CPU 2.15
Feature selection/transformation CPU 0.56
CPU-to-GPU transfer CPU → GPU 0.82
EfficientNet-B0 feature extraction GPU 2.41
Long-range attention GPU 0.53
Fuzzy feature weighting and fusion GPU 0.27
ShCNN classification GPU 0.58
Post-processing/output generation CPU 0.96
Neural-network inference subtotal GPU 3.79
Complete raw-image-to-prediction latency CPU + GPU 14.64
Table 13. Training and BTOA Hyperparameter-
Optimization Cost
Measure Value
BTOA Population Size 20
BTOA Iterations 30
Candidate Evaluations 600
Mean Candidate Evaluation
Time
52.2 s
BTOA Optimization Time 8.70 h
Final Model Training Time 46.8 min
Five Independent Final
Runs
3.90 h
Hardware NVIDIA RTX 3090 /
Tesla T4
Table 14. Friedman Repeated-Measures Analysis of
Macro-F1 across Five Matched Training Seeds
Architecture Mean Macro -
F1 (%)
Mean
Rank
CNN 91.72 1.00
ResNet-50 94.20 2.00
EfficientNet-B0 95.85 3.00
EfficientNet-B0 +
Attention
97.05 4.00
Proposed EfficientNet –
Fuzzy ShCNN
98.55 5.00
Friedman 𝜒𝐹2 20.00 —
Degrees of freedom 4 —
Asymptotic 𝑝-value 0.00050 —
Kendall's 𝑊 1.000 —
Received: September 4, 2026. Revised: October 3, 2026. 388
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
among the evaluated architectures, while Kendall’s
𝑊 = 1.000 reflects complete consistency in their
rank ordering across the five evaluated seeds.
To further examine the paired differences,
planned post-hoc comparisons between the proposed
framework and each baseline were conducted using
exact two -sided Wilcoxon signed -rank tests, with
Holm adjustment for multiple comparisons. The
proposed framework achieved a higher Macro -F1
score than each baseline across all five matched
training seeds. However, with only five non -zero
paired observations, the smallest attainable exact
two-sided Wilcoxon 𝑝 -value is 2/25 = 0.0625 .
Accordingly, each comparison produced an exact raw
𝑝 = 0.0625 , and none reached the conventional
significance threshold of 𝛼 = 0.05after multiplicity
adjustment. This result does not contradict the
significant omnibus Friedman test: the Friedman test
evaluates the overall systematic rank differences
across all five architectures, whereas each post -hoc
comparison contains only five pai red observations
and consequently has limited inferential resolution.
As shown in Table 14A, the proposed
EfficientNet–Fuzzy ShCNN framework achieved a
higher Macro-F1 score than each baseline across all
five matched training seeds. Nevertheless, none of
the individual pairwise comparisons reached
statistical significance at 𝛼 = 0.05 after Holm
adjustment. This outcome reflects the limited
inferential resolution of an exact two-sided Wilcoxon
signed-rank test with only five non -zero paired
observations. Therefore, the statistically significant
omnibus Friedman result should be interpre ted as
evidence of an overall systematic difference among
the evaluated architectures rather than as evidence
that every individual proposed -versus-baseline
comparison is statistically significant.
Table 15 presents the class -wise sensitivity and
Expected Calibration Error (ECE) obtained from the
same representative -run predictions on the
independent 210-image test set used for Table 5 and
Fig. 4. Consequently, the sensitivity values
correspond direc tly to the class -wise recall values
reported in Table 5: 98.11% for Healthy, 98.08% for
Bacterial Blight, 100.00% for Leaf Curl Virus, and
98.08% for Fusarium Wilt. The corresponding 95%
confidence intervals characterize the uncertainty
associated with these class-wise sensitivity estimates
on the representative test set. The calibration results
further show low class -wise ECE values, ranging
from 1.20% to 1.45%, indicating that the predicted
confidence values were closely aligned with the
observed class-wise correctness rates.
4.6 Explainability analysis
To improve the interpretability and transparency
of the proposed EfficientNet –Fuzzy ShCNN
framework, three complementary Class Activation
Mapping (CAM) techniques, namely Gradient -
weighted Class Activation Mapping (Grad -CAM),
Score-CAM, and Eigen -CAM, were employed as
qualitative interpretability tools to visualize the
image regions contributing to disease classification.
Because lesion -level ground -truth masks or
bounding-box annotations are unavailable for the
evaluated dataset, these activation maps are not
interpreted as quantitative evidence of lesion -
localization accuracy. Instead, they provide
qualitative insight into the spatial evidence utilized
by the proposed classifier when generating disease
predictions.
Fig. 6(a) illustrates representative Grad -CAM
visualizations generated by the proposed
EfficientNet–Fuzzy ShCNN framework. For the
Table 14A. Exact Paired Post-hoc Comparisons between the Proposed Framework and Baselines
Comparison Proposed
Higher
Wilcoxon
𝑊
Exact Two -
Sided 𝑝
Holm-
Adjusted 𝑝
Decision ( 𝛼 =
0.05)
Proposed vs CNN 5/5 seeds 0 0.0625 0.2500 Not significant
Proposed vs ResNet-50 5/5 seeds 0 0.0625 0.2500 Not significant
Proposed vs EfficientNet-B0 5/5 seeds 0 0.0625 0.2500 Not significant
Proposed vs EfficientNet -B0 +
Attention
5/5 seeds 0 0.0625 0.2500 Not significant
Table 15. Representative-Run Per-Class Sensitivity with 95% Bootstrap Confidence Intervals and Class-Wise ECE
Dataset Class Sensitivity (%) 95% Bootstrap CI ECE (%)
Cotton Dataset Healthy 98.11 [94.34, 100.00] 1.20
Cotton Dataset Bacterial Blight 98.08 [94.23, 100.00] 1.45
Cotton Dataset Leaf Curl Virus 100.00 [100.00, 100.00] 1.30
Cotton Dataset Fusarium Wilt 98.08 [94.23, 100.00] 1.38
Received: September 4, 2026. Revised: October 3, 2026. 389
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
evaluated samples, the activation patterns indicate
image regions that contributed strongly to the
corresponding class predictions. Disease -class
samples show concentrated responses over visually
distinctive portions of the leaves, whereas the healthy
sample exhibits a more distributed activation pattern.
Fig. 6(b) presents Score -CAM visualizations
obtained from the same representative test samples.
Score-CAM estimates the contribution of activation
maps through changes in the target -class score and
provides an alternative qualitative view of the spatial
evidence used for classification. The resulting maps
show relatively concentrated activation patterns over
visually informative leaf regions.
Fig. 6(c) depicts the corresponding Eigen -CAM
visualizations. Eigen -CAM derives class -activation
representations from the principal components of
convolutional feature maps and therefore provides a
complementary visualization of dominant spatial
feature response s. The generated maps indicate
regions receiving comparatively strong model
activation for the representative samples. Together,
Grad-CAM, Score -CAM, and Eigen -CAM provide
complementary qualitative perspectives on the spatial
evidence exploited by the proposed framework.
4.7 Uncertainty estimation
Predictive uncertainty and calibration were
evaluated to characterize the confidence associated
with the model predictions. Predictive entropy was
used as an uncertainty measure, with lower entropy
indicating a more concentrated and confident
predicted probability distribution. Expected
Calibration Error (ECE) was additionally used to
quantify the agreement between predicted confidence
and observed prediction accuracy.
For ECE computation, the confidence interval
[0, 1] was divided into 𝑀 = 15 equal-width bins. For
each test image 𝑖, prediction confidence was defined
as the maximum softmax probability, max𝑐𝑝𝑖,𝑐 ,
associated with the predicted class. For class -wise
ECE, test predictions were grouped according to their
ground-truth class, and ECE was calculated
separately for each class using the same 𝑀 =
15 equal-width confidence bins and maximum -
softmax-probability confidence definition. Empty
bins were excluded from the corresponding
summation. The same calibration procedure was
applied consistently across the evaluated runs.
A non-parametric bootstrap procedure was used
separately to estimate 95% confidence intervals for
class-wise sensitivity. Test -set predictions were
resampled with replacement at the image level using
1,000 bootstrap replicates. Class-wise sensitivity was
recomputed for each replicate, and the 2.5th and
97.5th percentiles of the resulting bootstrap
distribution were used as the lower and upper limits
of the 95% confidence interval.
Figure. 6(a) Grad-CAM visualization of the proposed EfficientNet–Fuzzy ShCNN framework
Received: September 4, 2026. Revised: October 3, 2026. 390
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
Figure. 6(b) Score-CAM visualization of the proposed EfficientNet–Fuzzy ShCNN framework
Figure. 6(c) Eigen-CAM visualization of the proposed EfficientNet–Fuzzy ShCNN framework
As shown in Table 16, the representative test run
yielded an overall mean predictive entropy of 0.136
and an overall ECE of 1.33%. The proposed
framework achieved a mean ECE of 1.28% across the
five independent training runs, as reported in Table 7.
These values correspond to different aggregation
levels: the 1.28% value in Table 7 represents the
mean ECE across five training runs, whereas the
1.33% value in Table 16 corresponds specifically to
the representative test run used for the detailed class-
wise pe rformance, uncertainty, and calibration
analyses.
Received: September 4, 2026. Revised: October 3, 2026. 391
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
Table 16. Representative-Run Predictive Uncertainty and
Calibration Results
Class Mean Predictive Entropy ECE (%)
Healthy 0.118 1.20
Bacterial
Blight 0.164 1.45
Leaf Curl
Virus 0.109 1.30
Fusarium Wilt 0.151 1.38
Overall 0.136 1.33
- Conclusion
The proposed EfficientNet –Fuzzy ShCNN
framework provides an effective approach for multi-
class cotton leaf disease classification under the
image variations represented in the evaluated
benchmark dataset. The framework integrates
EfficientNet-based deep feature extraction, long -
range attentio n, adaptive fuzzy feature weighting,
multi-level feature fusion, and a lightweight ShCNN
classifier to improve classification performance
while maintaining computational efficiency.
Experimental evaluation demonstrated an accuracy
of 98.72%, with precision, recall, and F1-score values
exceeding 98% and an error rate of 1.28%. Under the
common experimental protocol, the proposed
framework achieved higher classification
performance than CNN, ResNet50, EfficientNet-B0,
and EfficientNet -B0 with Attention. Compo nent-
wise ablation analysis further demonstrated the
contribution of the individual preprocessing,
attention, fuzzy -weighting, feature -fusion, ShCNN,
and optimization stages.
Statistical validation using a Friedman repeated -
measures test identified an overall difference among
the evaluated architectures. Exact paired Wilcoxon
signed-rank post-hoc tests with Holm correction were
additionally performed; however, the individual
pairwise comparisons did not reach the conventional
0.05 significance level, reflecting the limited
inferential resolution associated with five matched
training runs. The proposed framework requires 4.87
million trainable parameters and 0.46 GFLOPs, with
an average neural-network inference time of 3.79 ms
per image under the evaluated GPU environment.
Further evaluation using independently collected
field datasets across different geographic regions,
cultivars, seasons, and acquisition devices is required
to assess cross -domain generalization and practical
deployment suitability. - Limitations and future scope
Because the present optimization analysis was
conducted using a single predefined training -
validation partition, the reported BTOA
improvement should be interpreted as validation -
partition-specific evidence under the evaluated
benchmark protocol. Although strict separation
between optimization and independent test
evaluation was maintained, the present experiment
does not establish that the same optimization gain
would be reproduced across alternative validation
partitions. Repeated -split or nested model -selection
experiments will therefore be considered in future
work to quantify the sensitivity of the optimization
procedure to validation-set composition. Despite the
strong performance of the proposed EfficientNet –
Fuzzy ShCNN framework, certain limitations should
be acknowledged. In the first step of the process, the
model is assessed using a single dataset developed by
Kaggle for cotton leaf disease. The dataset may not
be representative of the diversity in field conditions,
including extreme lighting, occl usion, and
background clutter. Second, although data
augmentation could improve generalization, the
absence of large-scale and multi-source training and
testing datasets might limit robustness of models
across different geographic locations and cultivation
systems. Third, the current approach relies solely on
image-based features and does not incorporate
additional contextual information such as
environmental or temporal data. One limitation of the
present study is that all experiments were conducted
using a workstation equipped with an NVIDIA RTX
3090/T4 GPU. Although the proposed architecture is
computationally lightweight, deployment on
dedicated edge platforms such as NVIDIA Jetson,
Raspberry Pi, or mobile devices was not
experimentally evaluated. Future work will
investigate hardware -specific optimization using
TensorRT and TensorFlow Lite and assess real -time
performance in practical agricultural environments.
Code and Data Availability
The dataset used in this study is publicly available
from its original provider. The implementation code,
trained checkpoints, split manifest, per -image
predictions, and associated experimental artifacts
form part of ongoing doctoral research and are
currently subject to institutional restrictions on public
dissemination. These materials therefore cannot
presently be deposited in a public repository. The
manuscript provides the experimental configuration,
preprocessing parameters, dataset -partitioning
procedure, random seeds, model settings, and
Received: September 4, 2026. Revised: October 3, 2026. 392
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
evaluation protocol required to describe the reported
experiments. Release of additional research artifacts
is subject to institutional approval.
Conflicts of Interest
The authors declare that they have no conflicts of
interest regarding the publication of this paper.
Author Contributions
Sunita Chalageri : Conceptualization and
methodology; Dr. Kavita K Patil: Investigation and
data collection; Dr. Manjula H Nebagiri: Validation
and supervision; Saba T: Literature review and data
curation; Naitik S T: Implementation and
experimentation; Sandeep Kumar: Data analysis and
interpretation; Shrikant J. Honade: Resources and
technical review; G. J. Sahani: Visualization and
results analysis; Syed Sumera Ali: Manuscript
writing, review, and editing; Sridhar S K: Project
supervision and manuscript review. All author s
reviewed and approved the final version of the
manuscript.
References
[1] G. Kar, P. K. Ghosh, A. Das, R. Saxena, K.
Banerjee, and D. Vijay, “Commercial crops
(Jute, cotton and sugarcane)”, in Trajectory of
75 Years of Indian Agriculture After
Independence, pp. 315 -342, 2023, doi:
10.1007/978-981-19-7997-2_14.
[2] J. A. Wani, S. Sharma, M. Muzamil, S. Ahmad,
S. Sharma, and S. Singh, “Machine learning and
deep learning based computational techniques
in automatic agricultural diseases detection:
Methodologies, applications, and challenges”,
Archives of Computational Methods in
Engineering, Vol. 29, No. 1, pp. 641-677, 2022,
doi: 10.1007/s11831-021-09588-5.
[3] T. Shi, Y. Liu, X. Zheng, R. Zhang, L. Sun, and
W. Zhang, “Recent advances in plant disease
severity assessment using convolutional neural
networks”, Scientific Reports, Vol. 13, No. 1, p.
2336, 2023, doi: 10.1038/s41598-023-29230-7.
[4] T. Nivethitha and C. Venkataramanan, “A
quantum-enhanced deep learning framework for
precise leaf disease detection in tomato and
cotton crops”, Iranian Journal of Science and
Technology, Transactions of Electrical
Engineering, Vol. 49, No. 2, pp. 845-858, 2025,
doi: 10.1007/s40998-025-00912-4.
[5] F. Hu, Y. Zhang, X. Li, H. Wang, and Z. Chen,
“Investigation of an efficient multi -class cotton
leaf disease detection algorithm that leverages
YOLOv11”, Sensors, Vol. 25, No. 14, p. 4432,
2025, doi: 10.3390/s25144432.
[6] K. S. Prasad, R. Kumar, S. Sharma, and P. Singh,
“MCCLDP: Multi class cotton leaf diseases
prediction and classification using deep learning
model”, In: Proc. of Asia-Pacific Conference on
Communications, pp. 1 -6, 2025, doi:
10.1109/APCI65531.2025.11137216.
[7] J. Amin, M. Sharif, M. Raza, A. Saba, and S. A.
Shad, “Explainable neural network for
classification of cotton leaf diseases”,
Agriculture, Vol. 12, No. 12, p. 2029, 2022, doi:
10.3390/agriculture12122029.
[8] S. S. K., T. Sardar, B. S. S., N. S. T., G.
Thahniyath, and K. Dilip, “Transformer -based
multimodal deep learning for early disease
detection in smart healthcare systems”, In: Proc.
of International Conference on Machine
Learning and Autonomous Systems , pp. 1290 -
1295, 2026, doi:
10.1109/ICMLAS67792.2026.11483899.
[9] D. Anil, S. Kumar, R. Sharma, and P. Verma,
“Disease detection and diagnosis on the leaves
using image processing”, In: Proc. of IEEE
Global Conference on Computing and Media
Technologies, pp. 1 -6, 2022, doi:
10.1109/GCAT55367.2022.9972010.
[10] N. Ahmed, R. Navya, and A. Ananthanarayanan,
“A lightweight denoising convolutional neural
network for on -device artifact suppression”,
Engineering, Technology & Applied Science
Research, Vol. 16, No. 1, pp. 31484 -31489,
2026, doi: 10.48084/etasr.15428.
[11] C. K. Rai and R. Pahuja, “Classification of
diseased cotton leaves and plants using
improved deep convolutional neural network”,
Multimedia Tools and Applications, Vol. 82, No.
16, pp. 24811 -24835, 2023, doi:
10.1007/s11042-023-14933-w.
[12] B. Arathi and U. N. Dulhare, “Classification of
cotton leaf diseases using transfer learning -
DenseNet-121”, in Smart Trends in Computing
and Communications , Vol. 612, pp. 345 -355,
2023, doi: 10.1007/978-981-19-9228-5_33.
[13] A. Kaur, R. Sharma, S. Chattopadhyay, and K.
Joshi, “A fine -tuned deep learning -based
VGG16 model for cotton leaf disease
classification”, In: Proc. of IEEE International
Conference for Innovation in Technology, pp. 1-
6, 2024, doi:
10.1109/I2CT61223.2024.10544051.
[14] B. N. Pandey, S. Kumar, A. Sharma, and R.
Singh, “Cotton leaf disease classification using
deep learning based novel approach”, In: Proc.
of International Conference on Digital
Received: September 4, 2026. Revised: October 3, 2026. 393
International Journal of Intelligent Engineering and Systems, Vol.19, No.12, 2026 DOI: 10.22266/ijies2026.1231.18
This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
License details: https://creativecommons.org/licenses/by-sa/4.0/
Technologies, pp. 1 -6, 2023, doi:
10.1109/ICDT57929.2023.10150884.
[15] V. Rajasekar, S. Kumar, P. Singh, and R.
Sharma, “Detection of cotton plant diseases
using deep transfer learning”, Journal of Mobile
Multimedia, Vol. 18, No. 2, pp. 245 -262, 2022,
doi: 10.13052/jmm1550-4646.1828.
[16] N. Ahmed, A. Ananthanarayanan, R. Navya, S.
B. S., and P. M. S., “Image artifact removal
using DnCNN in low bandwidth wireless
communication”, In: Proc. of IEEE
International Conference for Women in
Innovation, Technology & Entrepreneurship, pp.
1-6, 2025, doi:
10.1109/ICWITE64848.2025.11307019.
[17] S. D., S. S. G., G. F. I., R. Navya, S. K. G., and
L. Anuj, “Pharmaceutical quality assurance via
Transformer-based deep learning models for
automated tablet defect detection”, In: Proc. of
International Conference on Artificial
Intelligence and Knowledge Discovery in
Concurrent Engineering , pp. 1 -6, 2025, doi:
10.1109/ICECONF65644.2025.11379551.
[18] S. S. K., S. Mathapati, S. L. H., S. B., S. B. S.,
and S. G. R., “AR -CropShield: A lightweight
offline mobile system for rice disease detection
and confidence-aware treatment guidance”, In:
Proc. of International Conference on Inventive
Material Science and Applications , pp. 1624 -
1630, 2026, doi:
10.1109/ICIMA68728.2026.11564261.
[19] A. Kaur, R. Sharma, S. Chattopadhyay, and K.
Joshi, “Cotton leaf disease classification using
fine-tuned VGG16 deep learning model”, In:
Proc. of World Conference on Communication
& Computing , pp. 1 -4, 2024, doi:
10.1109/WCONF61366.2024.10692185.
[20] N. Inamdar, M. Managuli, R. Koti, J. Jakati, S.
P. H., and P. Kulkarni, “Precision cotton disease
detection via transformer models applied to leaf
imagery”, Frontiers in Artificial Intelligence ,
Vol. 8, p. 1743264, 2026, doi:
10.3389/frai.2025.1743264.
[21] S. Naveena and K. Kavitha, “Gossypium
herbaceum: Folium disease identification and
classification using Efficient Net -Coordinate
Convolutional Neural Network (EcoNet)”,
Engineering Applications of Artificial
Intelligence, Vol. 152, p. 110701, 2025, doi:
10.1016/j.engappai.2025.110701.
论文原文与出版信息
International journal of intelligent engineering and systems