论文信息

Md Rifat Sarker / Jonayed Al-Faruk / Kaniz Fatema 等 6 人

PLoS ONE

论文解读

AI 辅助整理
方法
深度可分离卷积四阶段网络,结合BN、GAP、Dropout与Grad-CAM可视
发现
仅8万参数、0.31MB即达98.14%准确率,较DenseNet121参数少90倍。
机会点
可探索将该轻量模型部署到移动端、无人机与物联网边缘设备并做田间实时验证。

论文摘要

2024年全球茶叶产值达256亿美元,但叶部病害每年造成约20%的产量损失。现有基于深度学习的检测系统因计算需求过高且可解释性有限,在实际部署中仍不实用。本文提出一种专为四类茶叶病害分类(褐斑病、红锈病、红蜘蛛螨和健康叶片)设计的轻量级卷积神经网络架构,在分类精度与计算效率之间实现了优异平衡。该模型在四个渐进式特征提取模块中采用深度可分离卷积,并结合批归一化、全局平均池化和Dropout正则化,以最小化参数量同时保持判别性特征学习能力。在TeaLeafNet数据集(四类共5,000张均衡图像)上评估,该模型以仅80,959个可训练参数实现98.14%的分类准确率——相比DenseNet121(730万参数)减少90倍——模型大小仅为0.31 MB,仅需1.9129亿次浮点运算,约为VGG16(307.1亿次浮点运算)的1/160。模型展现出稳健的各类别性能,所有类别的F1分数均达0.97以上,其中两个类别的ROC-AUC值达到完美的1.000。通过定量插入(AUC:0.972)和删除(AUC:0.181)忠实度评估验证的Grad-CAM可视化确认,模型的预测由具有诊断相关性的叶片区域驱动,实现了农业应用所必需的模型透明度。与九种先进架构(包括DenseNet121、MobileNetV2、ResNet50和EfficientNetB1)的对比分析表明,尽管DenseNet121实现了略高的准确率(99.20%),但其参数量多90倍、计算操作多30倍,不适合边缘部署。所提出的框架在所有评估模型中提供了最紧凑的架构并具有竞争力的准确率,为精准农业系统中资源受限的茶叶病害诊断建立了新基准。通过实现准确、可解释且计算高效的病害检测,本工作解决了AI驱动的作物监测在移动设备、无人机和物联网农业平台上实际部署的关键障碍。

全文

PLoS ONE

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 1 / 33

OPEN ACCESS
Citation: Sarker MR, Al-Faruk J, Fatema K,
Das J, Saju MS, Akhi OJ (2026) LightTeaNet:
SeparableConv-based lightweight CNN with
Grad-CAM explainability for tea leaf disease
identification. PLoS One 21(11): e0363128.
https://doi.org/10.1371/journal.pone.0363128
Editor: Wei Song, Northeastern University,
CHINA
Received: July 28, 2026
Accepted: November 4, 2026
Published: November 18, 2026
Copyright: © 2026 Sarker et al. This is an open
access article distributed under the terms of
the Creative Commons Attribution License,
which permits unrestricted use, distribution,
and reproduction in any medium, provided the
original author and source are credited.
Data availability statement: Data is available
at kaggle; https://doi.org/10.34740/kaggle/
dsv/11770641.
Funding: The author(s) received no specific
funding for this work.
RESEARCH ARTICLE
LightTeaNet: SeparableConv-based lightweight
CNN with Grad-CAM explainability for tea leaf
disease identification
Md Rifat Sarker1, Jonayed Al-Faruk 1*, Kaniz Fatema1, Joyonta Das2,
Md. Shahabuddin Saju3, Ozifatun Jannat Akhi1
1 Department of Computer Science and Engineering, Mymensingh Engineering College, Mymensingh,
Bangladesh, 2 Department of Electrical and Electronic Engineering, Mymensingh Engineering College,
Mymensingh, Bangladesh, 3 Department of Civil Engineering, Mymensingh Engineering College,
Mymensingh, Bangladesh

  • jonayedalfaruk211282@gmail.com
    Abstract
    Global tea production, valued at USD 25.6 billion in 2024, faces annual yield losses
    of approximately 20% due to foliar diseases, yet existing deep learning-based detec-
    tion systems remain impractical for real-world deployment due to prohibitive compu-
    tational requirements and limited interpretability. This paper presents a
    purpose-built lightweight convolutional neural network architecture for four-class tea
    leaf disease classification (Brown Blight, Red Rust, Red Spider Mite, and Healthy
    Leaves) that achieves an exceptional balance between classification accuracy
    and computational efficiency. The proposed model leverages depthwise separable
    convolutions across four progressive feature extraction blocks, integrated with Batch
    Normalization, Global Average Pooling, and Dropout regularization to minimize
    parameter count while maintaining discriminative feature learning. Evaluated on
    the TeaLeafNet dataset (5,000 balanced images across four categories), the model
    achieves 98.14% classification accuracy with only 80,959 trainable parameters—rep-
    resenting a 90-fold reduction compared to DenseNet121 (7.3M parameters)—and a
    model size of merely 0.31 MB, while requiring only 191.29 million FLOPs, approxi-
    mately 160 × fewer than VGG16 (30.71B FLOPs). The model demonstrates robust
    per-class performance with F1-scores of 0.97 or higher across all categories and
    perfect ROC-AUC values of 1.000 for two classes. Grad-CAM visualization, validated
    through quantitative Insertion (AUC: 0.972) and Deletion (AUC: 0.181) faithfulness
    evaluation, confirms that the model’s predictions are driven by diagnostically relevant
    leaf regions, achieving model transparency essential for agricultural adoption. Com-
    parative analysis against nine state-of-the-art architectures—including DenseNet121,
    MobileNetV2, ResNet50, and EfficientNetB1—reveals that while DenseNet121
    achieves marginally higher accuracy (99.20%), it requires 90 × more parameters and
    30 × more computational operations, rendering it unsuitable for edge deployment.

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 2 / 33
The proposed framework offers the most compact architecture among all evaluated
models with competitive accuracy, establishing a new benchmark for resource-
constrained tea disease diagnosis in precision agriculture systems. By enabling
accurate, interpretable, and computationally efficient disease detection, this work
addresses critical barriers to the practical deployment of AI-driven crop monitoring on
mobile devices, unmanned aerial vehicles, and IoT-enabled agricultural sensors.

  1. Introduction
    Tea (Camellia sinensis) is one of the most widely consumed beverages in the world,
    with global production reaching approximately 6.7 million metric tons in 2023 and
    the global tea market valued at USD 25.6 billion in 2024 [1]. The industry supports
    the livelihoods of an estimated 13 million people worldwide, with major production
    concentrated in China, India, Kenya, Sri Lanka, and Turkiye [2]. However, tea trees
    grow in warm, humid environments that are inherently conducive to the spread of
    fungal, bacterial, and pest-related diseases. Foliar diseases including blister blight,
    gray blight, and brown blight are particularly significant as they adversely affect the
    bud and the two youngest leaves, causing direct loss of harvestable shoots [3]. Tea
    leaf diseases cause a decrease of around 20% in tea leaf yield annually, resulting in
    significant financial losses for tea farmers [4]. Accurate, timely, and scalable disease
    detection is therefore a critical requirement for the sustainable management of tea
    plantations at a global scale.
    Traditionally, tea disease diagnosis has depended on manual inspection by trained
    agronomists or experienced farmers. This process is time-consuming, labor-
    intensive, and highly subjective, making large-scale standardized monitoring and
    early warning difficult, especially in large plantations where the optimal treatment
    window is easily missed [5]. The absence of consistency in manual procedures has
    been highlighted by comparative research showing that human inspection accuracy
    for disease detection can differ by up to 30% between specialists and non-experts
    [6]. These limitations have motivated growing interest in automated, computer vision-
    based methods capable of delivering consistent, scalable, and objective diagnostic
    outputs without dependence on human intervention [7].
    Deep learning-based plant disease detection offers considerable progress com-
    pared to classical methods. When trained with large and high-quality datasets,
    these technologies can robustly detect diseases on plant leaves in early stages [8].
    Convolutional Neural Networks (CNNs) in particular have emerged as the domi-
    nant paradigm for image-based disease classification, owing to their ability to learn
    hierarchical feature representations directly from raw image data. CNN architectures
    are optimized from multiple perspectives including model structure and training label
    optimization, with models such as ResNet-50, InceptionV3, and MobileNetV2 being
    widely applied for plant disease detection [9]. Despite this strong classification per-
    formance, most standard CNN architectures are computationally expensive. State-of-
    the-art architectures such as VGG, ResNet, DenseNet, and EfficientNet achieve high
    Competing interests: The authors have
    declared that no competing interests exist.

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 3 / 33
accuracy by leveraging deep and complex network designs; however, these models typically involve millions of param-
eters and high computational complexity, making them unsuitable for deployment on mobile devices, unmanned aerial
vehicles (UAVs), and edge-based agricultural systems [10].
A comprehensive survey on deploying machine learning models at the network edge, emphasizing techniques such
as model compression, lightweight architectures, quantization, and pruning to address the resource constraints of edge
devices. The survey highlighted that efficient deep learning models can significantly reduce latency, communication over-
head, and energy consumption while maintaining satisfactory prediction performance. These findings demonstrate the
importance of lightweight and optimized deep learning architectures for real-time intelligent applications on resource-
constrained devices [11]. Beyond parameter reduction, architectural choices such as Global Average Pooling and Batch
Normalization have been shown to further stabilize training and reduce model size while improving generalization. Recent
lightweight models combining residual learning with efficient feature extractors have demonstrated robust performance,
achieving above 97% accuracy on large-scale plant disease datasets with as few as 3.51 million parameters, establishing
their suitability for mobile applications [12].
Despite these advances, the design of a purpose-built lightweight CNN tailored specifically for fine-grained, multi-
class tea leaf disease classification remains an underexplored problem. Key challenges persist, including limited
generalizability, small and imbalanced datasets, and poor real-world performance, with a focus needed on robust,
lightweight architectures suitable for edge deployment [ 13]. Furthermore, while model accuracy has been widely
reported in the literature, few studies provide a comprehensive evaluation of computational efficiency — including
FLOPs, parameter count, model size, and training time — alongside interpretability evidence, both of which are
essential for practical deployment. To address the need for model transparency, Grad-CAM uses the gradients of any
target concept flowing into the final convolutional layer to produce a coarse localization map highlighting important
regions in the image for predicting that concept, offering visual explanations for CNN decisions without requiring
architectural changes or retraining [ 14].
To address these gaps, this paper proposes a lightweight custom CNN model for four-class tea leaf disease classifica-
tion. The proposed architecture employs depthwise separable convolutions (SeparableConv2D) across four progressive
feature extraction blocks, combined with Batch Normalization for training stability, Global Average Pooling to substantially
reduce the parameter count, and Dropout regularization to suppress overfitting. The model is optimized using the Adam
optimizer with adaptive learning rate scheduling and early stopping. To ensure rigorous evaluation, we report classification
performance metrics including accuracy, precision, recall, F1-score, and multi-class ROC-AUC, alongside a full compu-
tational cost profile encompassing trainable parameter count, model size in megabytes, FLOPs, and total training time.
Model interpretability is further validated through Grad-CAM visualization combined with quantitative Insertion and Dele-
tion faithfulness evaluation.
The main contributions of this paper are as follows:

  1. A novel lightweight CNN architecture is introduced by integrating depthwise separable convolutions with Global Aver-
    age Pooling, enabling efficient feature extraction while substantially reducing model complexity without compromising
    classification accuracy.
  2. An extensive computational efficiency analysis is conducted by evaluating model parameters, FLOPs, memory con-
    sumption, model size, and execution time, highlighting the proposed model’s effectiveness for deployment in computa-
    tionally constrained environments.
  3. The reliability and transparency of the proposed framework are investigated through a comprehensive explainability
    analysis using Grad-CAM alongside quantitative Insertion and Deletion faithfulness metrics, confirming that model pre-
    dictions are guided by disease-relevant visual patterns.

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 4 / 33
4) To demonstrate real-world applicability, the proposed model is integrated into a lightweight web-based diagnostic plat-
form that enables real-time tea leaf disease recognition from uploaded images, facilitating practical adoption in smart
agriculture and field-based crop health monitoring.
2. Literature review
To begin, Liang et al. (2025) [15] proposed LTDDN, a lightweight deep learning framework for tea disease classifica-
tion based on EfficientNet-Lite0. The model incorporates a Channel Focus Attention (CFA) mechanism, Feature Reuse
Module (FRM), and Feature Enhancement Mobile Inverted Bottleneck Convolution Module (FEMBCM) to improve feature
extraction while reducing computational complexity. Experimental results showed that LTDDN achieved 98.03% accu-
racy on a tea disease dataset with only 2.998M parameters and 394.3M FLOPs, outperforming several CNN and Vision
Transformer models. The model also demonstrated strong generalization performance on PlantVillage and Mini-ImageNet
datasets and was successfully deployed on mobile devices for real-time disease recognition.
In a related study, Balasundaram et al. (2025) [16] proposed a tea leaf disease detection framework using a custom
CNN combined with MLP, SVM, and Decision Tree classifiers on a dataset of 5,867 tea leaf images across six disease
classes. The study employed preprocessing techniques including ROI extraction, HSV-based leaf segmentation, Canny
edge detection, leaf cropping, and Segment Anything Model (SAM)-based disease segmentation, achieving a maximum
accuracy of 95.06%. However, the model showed limitations due to SAM’s sensitivity to lighting variations, confusion
between visually similar diseases, and dataset class imbalance.
Moving toward detection-based approaches, Wang et al. (2024) [ 17] proposed YOLOv8-RCAA, a lightweight tea
leaf disease detection framework based on YOLOv8. The study utilized 995 tea leaf images from six disease cate -
gories and expanded the dataset to 4001 images through data augmentation techniques, including Gaussian blur,
random noise, hue adjustment, exposure enhancement, and random splicing. By integrating a RepVGG backbone,
CBAM attention mechanism, anchor-free detection head, and ATSS sample assignment strategy, the proposed
model achieved 98.14% mAP with only 0.81M parameters. However, the model was evaluated using data col -
lected from a single geographical region, which may limit its generalization capability under diverse environmental
conditions.
Building on this lightweight design philosophy, Yang et al. (2025) [18] proposed WaveLiteNet, a lightweight tea leaf
disease classification framework that integrates 2D discrete wavelet transform (DWT), MobileNetV3, CBAM attention, and
Focal Loss to enhance feature extraction under complex field conditions. The model was trained on 3,260 tea leaf images
from five disease categories, augmented through blurring, scaling, rotation, illumination adjustment, and noise injection,
achieving 98.70% classification accuracy with only 3.16 million parameters. However, the study focused on classification
rather than disease localization and was evaluated using data collected from a limited geographical region, which may
restrict its generalization capability.
In contrast to deep learning-only pipelines, Heng et al. (2024) [19] introduced a tea leaf disease classification frame-
work that combines a Hybrid Pooling CNN for feature extraction with a Weighted Random Forest (WRF) classifier opti-
mized using the Cuckoo Search Optimization (CSO) algorithm. The method was evaluated on the Tea Sickness Dataset
(TSD) containing 885 images of 7 tea leaf diseases, achieving 92.47% accuracy, 92.35% recall, and 92.26% preci-
sion. Images were preprocessed by converting RGB to HS color space and resizing to 150 × 150 pixels; however, the
approach is limited by the small dataset size, exclusion of healthy leaves, and increased training time due to WRF weight
optimization.
Similarly working with field-collected data, Rahman et al. (2024) [20] developed a CNN-based framework for automated
tea leaf disease detection using a primary dataset of 3,330 field-collected images from Bangladesh. The model classified
four categories—brown blight, grey blight, red rust, and healthy leaves—achieving an overall accuracy of 96.65%. Data
preprocessing included image resizing to 224 × 224 pixels, RGB conversion, normalization, and augmentation through

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 5 / 33
zooming, shearing, and flipping. Despite its high performance, the study was limited to four classes and relied on a cus-
tom CNN architecture without exploring advanced transfer-learning models.
Meanwhile, addressing data scarcity through generative augmentation, Li and Liao (2025) [21] proposed a tea dis-
ease recognition framework that integrates two-stage lesion segmentation, IC-GAN-based data augmentation, and an
Inception-based Dilated Convolutional Neural Network (IDCNN). The study utilized 400 field-collected tea leaf images
representing three disease classes and healthy leaves. After Graph-Cut and SVM-based lesion segmentation, IC-GAN
expanded the dataset to 8,000 images, enabling the IDCNN model to achieve 96.17% classification accuracy. However,
the authors reported that the effectiveness of IC-GAN decreases on larger datasets and that practical deployment remains
challenging under diverse field conditions.
More recently, extending detection to aerial imagery, Zhang et al. (2025) [ 22] proposed FCHE-YOLO, an enhanced
YOLO11-based framework for UAV-assisted tea leaf disease detection. The study utilized a self-constructed dataset
comprising 5,409 images generated from 2,432 high-quality UAV-acquired tea leaf images through data augmentation
techniques, including brightness adjustment, rotation, flipping, and random cropping. By integrating the FC_C3K2,
HSFPN, and Efficient-Head modules, the proposed model achieved a 98.1% mAP while maintaining a lightweight
architecture with 1.46M parameters and 4.2 GFLOPs. However, the model was validated on a single-location data -
set, and its robustness under diverse environmental conditions and large-scale tea plantations remains to be further
investigated.
In a related effort, Rahat et al. (2025) [23] developed a custom residual CNN for tea leaf disease classification using
4,000 images collected from Bangladesh tea gardens. After applying image augmentation and normalization techniques,
the model achieved 99.0% accuracy in detecting eight tea diseases and healthy leaves. Nevertheless, the model’s
performance depends on dataset diversity and requires further validation under varying environmental conditions and
resource-constrained deployment scenarios.
To improve interpretability alongside accuracy, Ozturk et al. (2025) [24] proposed an interpretable ensemble learning
framework for tea leaf disease classification by combining ResNet50, DenseNet121, EfficientNetB0, and MobileNetV2
with transfer learning and Grad-CAM visualization. Using an augmented Tea Sickness Dataset comprising 885 images
across eight classes, the model achieved 96% classification accuracy and demonstrated improved robustness and inter-
pretability. However, the study focused solely on image-level classification and was evaluated on a relatively small and
imbalanced dataset, limiting its applicability to disease localization tasks and large-scale field conditions.
Building on transformer-based detection architectures, Zhang et al. (2025) [25] proposed WMC-RTDETR, a lightweight
tea disease detection model based on RT-DETR that integrates Wavelet Transform Convolution (WTConv), Multiscale
Multihead Self-Attention (M2SA), and a Context-Guided Spatial Feature Reconstruction Feature Pyramid Network
(CSRFPN) to enhance multi-scale feature extraction and detection performance. Using an augmented dataset of 1,280
images derived from 160 original samples, the model achieved 97.7% mAP50 while reducing model parameters and com-
putational complexity. However, its evaluation was conducted on a relatively small, location-specific dataset, which may
limit generalization under diverse environmental conditions.
In another study focusing on channel-wise attention, Li and Zhao (2025) [26] proposed an ECA-ResNet50 model for
tea disease classification by integrating an Efficient Channel Attention (ECA) mechanism into ResNet50 and replacing the
initial 7 × 7 convolution with three 3 × 3 convolutions. The study used a tea disease dataset containing 885 images across
seven disease classes and healthy leaves, augmented through flipping, rotation, cropping, color transformation, and blur-
ring. The model achieved 93.06% accuracy, although its performance was constrained by the limited dataset size and the
similarity among multiple disease classes.
Furthermore, exploring cross-domain transfer learning, Yao et al. (2024) [27] proposed a transfer-learning-based
YOLOv7 framework enhanced with Triplet Attention, TSCODE, and NWD loss for tea leaf blight detection. Using an apple
leaf disease dataset (673 images) as the source domain and a tea leaf blight dataset (182 images) as the target domain,

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 6 / 33
augmented to over 3,000 samples, the model improved small-lesion detection in complex field environments, although its
effectiveness depended on limited annotated data and source-domain similarity.
Similarly leveraging transfer learning for broader pest coverage, Li et al. (2024) [28] introduced a lightweight
MobileNetV3-CA model for tea leaf disease and insect classification, combining Coordinate Attention and transfer learn-
ing to enhance feature extraction. The method utilized 22,380 images from 17 disease and pest categories, augmented
to 34,000 images and standardized to 224 × 224 pixels, achieving 95.88% accuracy on the test set. Despite its strong
performance, the study acknowledged the need for more diverse and early-stage pest images to improve real-world
applicability.
In a related approach to lightweight architecture design, Li et al. (2024) [29] proposed a lightweight I-MobileNetV2
framework for tea leaf disease identification by integrating a Coordinate Attention (CA) module, a Multi-branch Parallel
Convolution (MPC) module, and AMC-based model pruning. The study utilized a self-collected dataset of 2,544 images
covering five tea diseases and healthy leaves, which was expanded to 12,720 images through brightness adjustment,
chromaticity enhancement, rotation, and Gaussian noise augmentation. The proposed model achieved 96.12% accuracy
while reducing model parameters by approximately 40%, although misclassification remained challenging for visually simi-
lar diseases such as tea bud blight and tea leaf blight.
Likewise, using a large-scale Kaggle dataset, Hairah et al. (2024) [30] presented a MobileNetV2-based CNN framework
for tea leaf disease classification using a Kaggle dataset of 5,980 images across six classes. The images were resized to
224 × 224 pixels, normalized, and augmented through rotation, shearing, and flipping operations. The proposed approach
achieved 94.55% classification accuracy while maintaining low computational complexity, although its evaluation on
white-background images may limit real-world field applicability.
To improve the detection of small and visually complex tea leaf diseases, Zhan et al. (2024) [31] developed BHC-
YOLO based on YOLOv8n, incorporating BiFormer attention for target localization, Haar wavelet-based downsampling
for feature preservation, and a lightweight feature fusion module. The model was trained on 4,000 field-acquired images
augmented to 10,000 samples and demonstrated superior performance compared with YOLOv8n, YOLOv9s, and YOLO-
Tea, achieving 92.2% precision and 94.5% mAP@0.5. However, the framework requires high-quality images and exhibits
reduced accuracy under poor illumination conditions.
In a complementary effort to improve lightweight tea pest and disease detection, Chen et al. (2024) [32] introduced
TeaViTNet based on MobileViT, incorporating PANet for multi-scale feature fusion, RFBNet for receptive field enhance-
ment, EMA attention for contextual learning, and ODCSPLayer for richer feature representation. Images were resized
to 640 × 640 pixels and trained using the Adam optimizer for 300 epochs. Experimental results showed that TeaViTNet
achieved 89.1% mAP@0.5 while maintaining a compact model size of 4.5 M parameters, making it suitable for mobile and
edge computing applications.
In a comparative study of detection backbones, Deng et al. (2024) [33] presented an automated tea leaf disease and
pest detection system based on comparative evaluation of YOLOv8s, YOLOv9s, and YOLOv10s object detection models.
The study employed a self-collected dataset containing tea leaf images of brown blight, gray blight, and tea leafhopper,
with preprocessing involving image annotation, resizing, and dataset partitioning for training and validation. Experimen-
tal results showed that YOLOv10s achieved the best performance with an mAP@0.5 of 84.4% and superior processing
speed; however, the model exhibited reduced generalization due to region-specific data collection and limited representa-
tion of rare diseases and pests.
Along similar lines, Xia et al. (2024) [34] proposed an improved YOLOv7-based tea disease detection framework by
replacing the CSPDarkNet53 backbone with MobileNeXt, incorporating a Bi-level Routing Attention (BRA) mechanism,
and adopting the SIoU loss function to enhance feature extraction and localization accuracy. The study utilized a field-
collected dataset of 1,823 tea leaf images covering four diseases (anthracnose, leaf blight, bituminous coal disease, and
zonate spot), which was expanded to 4,280 images through color enhancement, flipping, random rotation, and contrast

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 7 / 33
augmentation. Experimental results demonstrated 92.1% mAP and 93.5% precision while reducing model parameters
by 34.6%; however, the model was evaluated on only four disease categories, indicating the need for larger and more
diverse datasets to improve generalization in complex tea garden environments.
To further improve tea leaf disease recognition under limited training data conditions, Kumar et al. (2024) [35] employed
K-means clustering for diseased-region extraction and EfficientNet with ImageNet-based transfer learning for classifica-
tion. Images were resized to 224 × 224 pixels and augmented using geometric transformations to mitigate overfitting and
increase sample diversity. Experimental results demonstrated superior performance compared with SVM, Random Forest,
CNN, and C-DCGAN models, achieving an overall accuracy of 91%; however, the framework was validated on only three
tea disease classes. Building on this, Leng et al. (2024) [36] introduced YOLO-DBD, an improved YOLOv8 framework that
integrates deformable convolutions, Bi-level Routing Attention, Dynamic Head, and Focal-CIoU loss for robust harmful
tea leaf detection in complex field environments. Evaluated on a self-collected dataset containing isolated, clustered, and
occluded harmful leaves, the model achieved 86.8% mAP@0.5 and demonstrated superior detection accuracy over sev-
eral state-of-the-art detectors. Nevertheless, the enhanced architecture increased computational complexity and reduced
inference speed relative to lightweight YOLO models.
To address challenges under low-light conditions, Ye et al. (2024) [37] introduced a low-light tea disease severity
detection framework by integrating an improved EnlightenGAN with a YOLOv8-ASFF architecture. Using a dataset of
3,743 annotated images of three tea diseases, the method employed low-light image enhancement, ResNet50-based
feature extraction, attention mechanisms (ISAM/ICAM), ASFF fusion, ODConv, and EIoU loss, achieving 95.26% mAP
and 87.47% detection accuracy. Despite its strong performance, the model was validated on a region-specific dataset and
may require further evaluation for broader generalization. Similarly targeting robustness against complex backgrounds,
Sun et al. (2024) [38] introduced TCSA, a CSPDarkNet53-based tea disease detector enhanced with self-attention, chan-
nel attention, and spatial attention modules. Trained on 3,676 augmented images from six disease categories, the model
achieved 94.6% mAP, demonstrating improved robustness against multiscale diseases, occlusions, and complex back-
grounds.Ye et al. (2024) [39] introduced YOLOv8-RMDA, a lightweight tea disease detection model that integrates RFC-
BAM, MixSPPF, Dynamic Head attention, AKConv, and Inner-EIoU loss to enhance small-target lesion detection. Using
3,357 annotated images of four tea diseases, the model achieved 93.04% mAP@0.5, surpassing several state-of-the-art
detectors.
Finally, recent studies have explored efficient deep learning approaches for complex image analysis. Zhao et al. [40]
proposed RGD-UNet for accurately segmenting clustered and overlapping signet-ring cells while maintaining computa-
tional efficiency. Similarly, Zhao et al. [41] introduced UADC, a semi-supervised segmentation framework that combines
multi-view learning and uncertainty-guided pseudo-label refinement to improve segmentation performance. Although
developed for medical imaging, these approaches highlight the importance of efficient feature extraction and robust learn-
ing strategies for complex visual recognition tasks.
3. Methodology
This study proposes a lightweight and computationally efficient Convolutional Neural Network (CNN) framework for auto-
mated tea leaf disease classification. The overall workflow consists of seven major stages: data collection, data prepro-
cessing, model development, training, evaluation, prediction, and explainability analysis using Gradient-weighted Class
Activation Mapping (Grad-CAM). The complete framework is illustrated in Fig 1, while the detailed architecture of the
proposed lightweight CNN is presented in Fig 2.
3.1 Overall framework
The proposed system follows an end-to-end deep learning pipeline designed to accurately identify tea leaf diseases from
RGB images. Initially, tea leaf images belonging to four categories — three disease classes (Brown Blight, Red Rust,

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 8 / 33
and Red Spider Mite) and one healthy class (Healthy Leaves) — are collected and organized into class-specific folders.
Subsequently, image preprocessing techniques, including resizing, normalization, and data augmentation, are applied to
improve model generalization and reduce overfitting.
The preprocessed images are then fed into a lightweight CNN architecture based on depthwise separable convo -
lutions. The extracted feature representations are learned through a series of convolutional blocks and subsequently
classified using fully connected layers. During training, the model parameters are optimized using the Adam opti -
mizer and categorical cross-entropy loss function. The trained model is evaluated using multiple performance met -
rics, including accuracy, precision, recall, F1-score, confusion matrix, and ROC analysis. Finally, Grad-CAM visual -
ization is employed to interpret the model’s predictions by highlighting the most discriminative disease regions within
the leaf images.
3.2 Dataset description and statistics
The experiments conducted in this study utilize the TeaLeafNet dataset [42], a publicly available benchmark dataset
hosted on Kaggle and specifically designed for automated tea leaf disease recognition. The dataset was developed
to support computer vision and deep learning research for early diagnosis of tea leaf diseases, thereby contributing to
Fig 1. Workflow of the proposed mode for tea leaf Disease Classification.
https://doi.org/10.1371/journal.pone.0363128.g001

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 9 / 33
precision agriculture and intelligent crop monitoring systems. Early and accurate disease identification can assist farmers
in implementing timely interventions, reducing crop losses, and improving tea production efficiency.
TeaLeafNet contains images from four categories representing three common tea leaf diseases and one healthy class:
Brown Blight (BB), Red Rust (RR), Red Spider Mite (RSM), and Healthy Leaves (GL). The images were collected under
real-world field conditions across different seasons. The dataset follows an acronym-based naming convention, where the
disease class is represented by abbreviations (e.g., BB, RR, and RSM), while the suffixes A, S, and W indicate the season
of image acquisition, corresponding to Autumn, Summer, and Winter, respectively. This seasonal variation introduces
diverse environmental conditions and symptom appearances, enhancing the robustness and generalization capability of
deep learning models.
The original images are stored in JPEG format and exhibit varying spatial resolutions, including dimensions such as
4926 × 2304 pixels and 2304 × 1286 pixels. All images are represented in the RGB color space, preserving important chro-
matic and textural characteristics associated with disease symptoms. To minimize background interference, the dataset
creators performed leaf segmentation and background removal during dataset preparation. As a result, each image con-
tains an isolated tea leaf positioned against a uniform black background, allowing classification models to focus primarily
on disease-related visual features.
Fig 2. Proposed Lightweight CNN Architecture.
https://doi.org/10.1371/journal.pone.0363128.g002

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 10 / 33
To address the class imbalance present in the original dataset, a balanced subset containing 5,000 images was uti-
lized in this study. Each class contributes 1,250 images, ensuring equal representation across all categories and reducing
potential bias during model training and evaluation. The class-wise distribution of the dataset is presented in Table 1.
Prior to training, all images were resized to 224 × 224 × 3 pixels to match the input requirements of the proposed deep
learning framework. The combination of balanced class representation, seasonal diversity, high-resolution image acqui-
sition, and disease-specific symptom variability makes TeaLeafNet a challenging and reliable benchmark for evaluating
automated tea leaf disease classification systems.
The dataset characteristics summarized in Table 2 demonstrate that TeaLeafNet provides a well-structured and bal-
anced collection of tea leaf images suitable for training and validating deep learning models. Its controlled background,
seasonal diversity, and equal class distribution enable a comprehensive assessment of disease classification performance
under realistic agricultural scenarios.
3.3 Data Partitioning
The dataset comprising 5,000 images was divided into training, validation, and testing subsets to facilitate model devel -
opment and evaluation. Specifically, 3,496 images (70%) were used for training, while 752 images (15%) each were
allocated for validation and testing. The data were distributed proportionally across all disease classes to maintain a bal -
anced class representation and ensure reliable model training and unbiased performance assessment. The class-wise
distribution of images across the three subsets is illustrated in Fig 3, while the overall partition statistics are summarized
in Table 3.
Table 1. Class-wise distribution of the TeaLeafNet dataset used in this study.
Class Abbreviation Number of Images
Brown Blight BB 1,250
Red Rust RR 1,250
Red Spider Mite RSM 1,250
Healthy Leaves GL 1,250
Total – 5,000
https://doi.org/10.1371/journal.pone.0363128.t001
Table 2. Summary of the TeaLeafNet Dataset Characteristics.
Dataset Property Value
Dataset Name TeaLeafNet
Source Kaggle
Total Images 5,000
Number of Classes 4
Images per Class 1,250
Image Format JPEG
Color Space RGB
Original Resolution Variable
Background Segmented black background
Input Size Used 224 × 224 × 3
https://doi.org/10.1371/journal.pone.0363128.t002

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 11 / 33
3.4 Data preprocessing and augmentation
3.4.1 Image resizing. Since deep learning models require fixed-dimensional inputs, all tea leaf images were resized
to 224 × 224 × 3 pixels. Standardizing image dimensions ensures computational consistency and facilitates efficient batch
processing during training.
3.4.2 Data augmentation. Although the TeaLeafNet dataset provides a balanced distribution of disease categories,
further data augmentation was employed to enhance model robustness, improve generalization capability, and reduce
the risk of overfitting during training. Data augmentation introduces controlled variations in the training samples, enabling
the network to learn disease-specific features that remain invariant under different imaging conditions and environmental
variations. Several online augmentation techniques were applied, including horizontal flipping, random rotation, zooming,
width shifting, height shifting, and brightness adjustment. These transformations simulate realistic variations in leaf
orientation, scale, illumination, and spatial positioning that may occur during field image acquisition. Consequently, the model
becomes more resilient to unseen conditions and better equipped to recognize disease symptoms across diverse scenarios.
As a result of the augmentation process, the total number of training images increased from the original 3,496 images
to 13,984 images, substantially enriching the diversity of the training dataset. This expanded dataset provides a wider
range of visual representations for each disease category, thereby facilitating more effective feature learning and improv-
ing the model’s generalization performance on unseen samples.
Representative examples of the augmented tea leaf images are presented in Fig 4, illustrating the diversity intro-
duced through the augmentation pipeline while preserving the underlying disease characteristics essential for accurate
classification.
Fig 3. Class-wise distribution of images across the training, validation, and testing.
https://doi.org/10.1371/journal.pone.0363128.g003
Table 3. Distribution of Images Across Dataset Partitions.
Dataset Split Number of Images Percentage (%)
Training 3,496 70
Validation 752 15
Testing 752 15
Total 5,000 100
https://doi.org/10.1371/journal.pone.0363128.t003

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 12 / 33
3.4.3 Image normalization. Pixel intensity values were normalized to the range [0,1] using:
Inorm = I255 , (1)
where (I) denotes the original pixel intensity and (I norm) represents the normalized pixel value. Normalization stabilizes
gradient propagation, accelerates convergence, and improves numerical stability during optimization.
Fig 4. Some samples of the augmented images.
https://doi.org/10.1371/journal.pone.0363128.g004

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 13 / 33
3.5 Proposed lightweight CNN architecture
3.5.1 Design motivation. Conventional convolutional neural networks often achieve high classification performance
at the expense of substantial computational complexity and memory requirements. Such architectures may not be suitable
for practical agricultural applications where computational resources are limited. To address this limitation, a lightweight
CNN architecture based on depthwise separable convolutions was developed.
Depthwise separable convolution decomposes a standard convolution into depthwise and pointwise operations,
significantly reducing the number of trainable parameters and floating-point operations (FLOPs) while preserving feature
extraction capability. This design enables efficient learning of disease-specific visual patterns without introducing exces-
sive computational overhead.
3.5.2 Feature extraction module. The proposed network consists of four sequential convolutional blocks. Each block
contains:
• SeparableConv2D layer (3 × 3 kernel)
• Batch Normalization layer
• MaxPooling2D layer (2 × 2)
The number of convolutional filters gradually increases across the network to enable hierarchical feature learning. Low-
level texture and edge information are captured in the initial layers, while high-level disease-specific representations are
learned in deeper layers.
The feature extraction configuration is summarized as follows:
Block 1: 32 filters → Output size: 112 × 112 × 32
Block 2: 64 filters → Output size: 56 × 56 × 64
Block 3: 128 filters → Output size: 28 × 28 × 128
Block 4: 256 filters → Output size: 14 × 14 × 256
Each convolution employs ReLU activation and same padding. Batch normalization is incorporated to reduce internal
covariate shift and improve training stability, whereas max pooling progressively reduces spatial dimensionality and com-
putational burden.
The operation of each feature extraction block can be represented as:
Fi = Pool
(
BN
(
ReLU
(
SeparableConv
(
Xi
))))
(2)
where (Xi) denotes the input feature map and (Fi) represents the corresponding output feature map.
3.5.3 Feature aggregation and classification head. Following the final convolutional block, a Global Average
Pooling (GAP) layer aggregates spatial feature information into a compact feature vector. Compared with traditional
flattening operations, GAP significantly reduces parameter count and helps prevent overfitting.
The aggregated feature vector is subsequently passed through a fully connected layer comprising 128 neurons with
ReLU activation. A dropout layer with a rate of 0.4 is then employed to further improve generalization by randomly deacti-
vating neurons during training.
Finally, a Softmax output layer with four neurons generates class probabilities corresponding to the four disease
categories.
The Softmax function is expressed as:

P(y = i)= ezi∑4
j=1 ezj (3)

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 14 / 33
where (zi) denotes the output score associated with class (i).
The predicted disease class is determined by selecting the category with the highest posterior probability.
3.6 Model training strategy
The proposed CNN was trained using the Adam optimization algorithm owing to its adaptive learning capability and effi-
cient convergence characteristics. Adam combines momentum-based optimization with adaptive learning rates, making it
particularly effective for deep learning applications.
Since tea leaf disease identification constitutes a multi-class classification problem, categorical cross-entropy was
selected as the objective function:

L =–
C∑
i=1
yilog
(ˆyi)
(4)
where (C) represents the number of disease classes, (yi) denotes the true label, and (ŷi) indicates the predicted class
probability.
During training, network weights were iteratively updated to minimize classification error while validation performance
was continuously monitored to ensure optimal generalization.
3.7 Architecture of the proposed custom lightweight CNN
The proposed model is a fully custom-designed lightweight Convolutional Neural Network built from scratch without relying
on pretrained ImageNet weights. The key architectural innovations are as follows.
The model is structured as a Sequential network consisting of four convolutional blocks followed by a classification
head. Each block employs SeparableConv2D instead of standard Conv2D, which decomposes the convolution into a
depthwise spatial convolution followed by a pointwise convolution. This decomposition drastically reduces the number of
parameters and floating-point operations compared to conventional convolution. Each block further incorporates Batch-
Normalization to stabilize and accelerate convergence, followed by MaxPooling2D (pool size 2 × 2) to progressively reduce
spatial dimensions.
The four blocks use filter sizes of 32, 64, 128, and 256, respectively, enabling the network to learn a hierarchical feature
representation from low-level edges to high-level semantic disease patterns. Following the convolutional blocks, Glo-
balAveragePooling2D is applied instead of Flatten, which significantly reduces the number of parameters by averaging
each feature map into a single value. A Dense layer with 128 neurons and ReLU activation, followed by a Dropout layer
(rate = 0.4) for regularization, precedes the final softmax output layer with 4 neurons corresponding to the four tea disease
classes. The architecture is summarized in Table 4.
Table 4. Architecture of the proposed Custom Lightweight CNN model.
Block Layer Type Filters/ Units Key Property
Block 1 SeparableConv2D + BN + MaxPool 32 Input: 224 × 224 × 3
Block 2 SeparableConv2D + BN + MaxPool 64 Spatial reduction
Block 3 SeparableConv2D + BN + MaxPool 128 Feature deepening
Block 4 SeparableConv2D + BN + MaxPool 256 High-level features
— GlobalAveragePooling2D — Parameter reduction
FC Dense (ReLU) + Dropout (0.4) 128 Regularization
Output Dense (Softmax) 4 Classification
https://doi.org/10.1371/journal.pone.0363128.t004

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 15 / 33
3.8 Evaluation metrics
In this paper, four standard classification metrics are used to evaluate model performance: Accuracy, Precision, Recall,
and F1-Score. These metrics are defined as follows [15]:
Accuracy = (TP + TN)/(TP + FP + TN + FN)
Precision = TP/(TP + FP)
Recall = TP/(TP + FN)
F1 – Score = 2 × (Precision × Recall)/(Precision + Recall)
where TP (True Positive) represents the number of samples correctly identified as positive, FP (False Positive) represents
the number of samples incorrectly identified as positive, TN (True Negative) represents the number of samples correctly
identified as negative, and FN (False Negative) represents the number of samples that are incorrectly identified as nega-
tive. All metrics are computed using weighted averaging across the four classes to account for any class imbalance.
In addition to classification performance, the computational complexity of each model is evaluated using the total num-
ber of trainable parameters (Params) and approximate model size in megabytes (MB). The parameter count of a convolu-
tional layer is formulated as [15]:
Params = K h × Kw × Cin × Cout (5)
where Kh and Kw represent the height and width of the convolution kernel, and Cin and Cout represent the number of input
and output channels, respectively. The approximate model size in MB is derived from Params as (Params × 4 bytes)/
(1024²). Training time in seconds is also recorded to assess practical efficiency.
3.9 Analysis of computational efficiency
The adoption of SeparableConv2D as the core building block provides the primary computational advantage in the pro-
posed model. The floating-point operations (FLOPs) of a convolutional layer are formulated as [15]:
FLOPs = C in × H × W × Kw × Kh × Cout (6)
Where H and W represent the height and width of the input feature map, Kh and Kw represent the convolution kernel
dimensions, and Cin and Cout represent the number of input and output channels, respectively.
In standard convolution, a layer with input channels Cin, output channels Cout, and kernel size k × k requires k2 × Cin×Cout
multiplications per spatial location. In depthwise separable convolution, this cost is reduced to k2 × Cin+Cin×Cout, represent-
ing a reduction factor of approximately 1/Cout+1/k2. For k = 3 and typical filter sizes, this yields a reduction of approximately
8–9× in computation per layer.
Furthermore, GlobalAveragePooling2D eliminates the large fully connected layers that are characteristic of classic CNN
architectures such as VGG16 (which use multiple dense layers of 4096 neurons), replacing them with a single averaged
feature per channel. This alone removes millions of parameters from the classification head while also serving as a strong
regularizer against overfitting.

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 16 / 33
The combination of these design choices results in a model that trains and performs inference significantly faster than
the transfer learning comparison models, as evidenced by the training time measurements recorded in the experiments.
3.10 Explainability analysis using Grad-CAM
To enhance the interpretability of the proposed lightweight CNN, Gradient-weighted Class Activation Mapping (Grad-CAM)
was employed as a post-hoc explainability technique. Grad-CAM generates visual heatmaps that highlight the image
regions contributing most significantly to the model’s predictions. By overlaying these heatmaps on the original tea leaf
images, it becomes possible to verify whether the model focuses on disease-affected areas rather than irrelevant fea-
tures. This analysis improves the transparency, reliability, and practical applicability of the proposed framework for tea leaf
disease diagnosis.
3.11 Novelty and contribution of the proposed framework
The proposed framework introduces a lightweight and interpretable CNN architecture specifically designed for tea leaf dis-
ease classification. Unlike conventional deep CNN models that require substantial computational resources, the proposed
architecture utilizes depthwise separable convolutions and global average pooling to significantly reduce model complexity
while maintaining effective feature extraction capability. Furthermore, the integration of Grad-CAM provides visual inter-
pretability, enabling transparent and trustworthy disease diagnosis. These characteristics make the proposed framework
suitable for practical deployment in intelligent agricultural monitoring systems and resource-constrained environments.
4. Results and analysis
4.1 Experimental setup and training parameters
All experiments are conducted on Google Colab utilizing an NVIDIA GPU environment. The dataset used in this study is
an augmented tea leaf disease dataset organized into train, validation, and test splits. The dataset comprises four catego-
ries: three disease classes — BB (Brown Blight), RR (Red Rust), and RSM (Red Spider Mite) — and one healthy class,
GL (Healthy Leaves). All input images are resized to a fixed resolution of 224 × 224 pixels and normalized to the [0, 1]
range by rescaling pixel values by a factor of 1/255.
For the proposed Custom Lightweight CNN model, the batch size is set to 32 and the maximum number of training
epochs is set to 50. The Adam optimizer is employed with an initial learning rate of 0.0001 and categorical cross-entropy
as the loss function. To prevent overfitting and ensure optimal convergence, two callbacks are applied during training:
EarlyStopping with a patience of 7 epochs (monitoring validation loss and restoring best weights), and ReduceLROnPla-
teau with a reduction factor of 0.5 and a patience of 3 epochs. These adaptive training strategies ensure that the model
converges efficiently without unnecessary computation.
For the nine comparison models — MobileNetV2, ResNet50, DenseNet121, InceptionV3, Xception, NASNetMo -
bile, VGG16, EfficientNetB1, and ResNet101 — all are trained using ImageNet pre-trained weights with their base
layers frozen (transfer learning). Each comparison model is trained for a fixed 50 epochs using the same Adam
optimizer (learning rate = 0.0001), batch size of 32, and categorical cross-entropy loss. A unified classification head
consisting of GlobalAveragePooling2D → Dense(256, ReLU) → Dropout(0.5) → Dense(4, Softmax) is appended to
each base model.
4.2 Comparison of model performance
The performance of the proposed lightweight CNN, LightTeaNet, is evaluated against nine widely used deep learning
architectures in terms of classification performance and computational complexity. All models in the initial comparison use
ImageNet-pretrained weights with frozen base layers and a unified classification head, trained for 50 epochs using the

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 17 / 33
Adam optimizer (learning rate of 0.0001) and categorical cross-entropy loss. The evaluation considers accuracy, preci-
sion, recall, F1-score, total parameters, model size, FLOPs, and training time.
As presented in Table 5, LightTeaNet achieves an accuracy of 98.14%, precision of 98.17%, recall of 98.14%, and
F1-score of 98.14%. DenseNet121 achieves the highest accuracy among these models at 99.20%, followed by Mobile-
NetV2 (98.67%) and InceptionV3 (98.54%). However, LightTeaNet offers substantial computational advantages, requiring
only 80,959 parameters, 0.31 MB of storage, and 191.29 million FLOPs. Its model size is approximately 32 times smaller
than MobileNetV2 and 90 times smaller than DenseNet121. Furthermore, LightTeaNet records the shortest training time of
6,820.02 s. These findings demonstrate the efficiency of the proposed architecture while maintaining competitive classifi-
cation performance.
The pretrained backbones expect different input conventions: raw [0, 255] values with internal normalization (Efficient-
Net), mean-subtracted BGR input (ResNet, VGG16), or scaling to [0, 1] or [− 1] with normalization (DenseNet121, Mobile-
NetV2, InceptionV3, Xception, NASNetMobile). The initial benchmark applied a uniform [0, 1] scaling to all models, which
did not match these conventions and degraded the pretrained features of EfficientNetB1, ResNet50, and ResNet101.
To further assess performance against selected pretrained architectures, Table 6 presents a comparison of LightTea-
Net with MobileNetV2, DenseNet121, and EfficientNetB1, including inference latency. DenseNet121 achieves the high-
est accuracy and F1-score of 100.00%, followed by EfficientNetB1 at 99.60% and MobileNetV2 at 99.34%. LightTeaNet
achieves 98.14% for both metrics while maintaining the lowest computational complexity, with 80,959 parameters, 191.29
million FLOPs, and a model size of only 0.31 MB. In comparison, the pretrained models require substantially more param-
eters and computational operations. Although inference latency is unavailable for LightTeaNet, its compact architecture
indicates potential for deployment on resource-constrained agricultural devices.
Table 5. Comparison of classification performance across all models on the tea leaf disease test set.
Model Accuracy Precision Recall F1-Score Total Parameters Model Size (MB) Total FLOPS Training Time (s)
DenseNet121 0.9920 0.9920 0.9920 0.9920 7,300,932 27.85 5700946776 8337.90
MobileNetV2 0.9867 0.9869 0.9867 0.9867 2,586,948 9.87 613383896 8829.84
InceptionV3 0.9854 0.9855 0.9854 0.9854 22,328,356 85.18 5694406872 8179.83
Xception 0.9801 0.9802 0.9801 0.9801 21,387,052 81.59 9135147736 8348.78
NASNetMobile 0.9774 0.9773 0.9774 0.9773 4,541,336 17.32 1146302052 8845.83
VGG16 0.9521 0.9525 0.9521 0.9521 14,847,044 56.64 30713219864 8345.05
ResNet50 0.8258 0.8227 0.8258 0.8184 24,113,284 91.98 7752248216 8914.08
ResNet101 0.7207 0.7189 0.7207 0.7207 43,183,748 164.73 15195613080 8323.62
EfficientNetB1 0.2500 0.0625 0.2500 0.1000 6,904,203 26.34 1181657035 8154.12
Our Model 0.9814 0.9817 0.9814 0.9814 80,959 0.31 191290328 6820.02
https://doi.org/10.1371/journal.pone.0363128.t005
Table 6. Comparative performance and computational complexity of LightTeaNet and end-to-end fine-tuned pretrained CNNs for tea leaf
disease classification.
Model Initialization Accuracy
(%)
F1-score
(%)
Parame-
ters
FLOPs Model
size (MB)
Inference latency
(ms/image)
MobileNetV2 ImageNet pretrained 99.34 99.34 2,586,948 613,383,896 9.87 86.47
DenseNet121 ImageNet pretrained 100.00 100.00 7,300,932 5,700,946,776 27.85 98.39
EfficientNetB1 ImageNet pretrained 99.60 99.60 6,904,203 1,181,657,035 26.34 87.28
LightTeaNet (proposed) Random initialization (from scratch) 98.14 98.14 80,959 191,290,328 0.31 N/A
https://doi.org/10.1371/journal.pone.0363128.t006

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 18 / 33
4.3 Per-class performance analysis
To provide a more detailed understanding of the model’s classification behavior, Table 7 presents the per-class precision,
recall, and F1-score for the proposed Custom Lightweight CNN, as computed by the classification report.
4.4 Confusion matrix analysis
Fig 5 presents the confusion matrix of the proposed Custom Lightweight CNN evaluated on the test set. The confusion
matrix provides a granular, class-level view of the model’s classification behavior, where the diagonal elements represent
correctly classified samples and the off-diagonal elements represent misclassifications. The rows correspond to the true
labels and the columns correspond to the predicted labels across the four categories: BB (Brown Blight), GL (Healthy
Leaves), RR (Red Rust), and RSM (Red Spider Mite).
Out of a total of 752 test samples, the model correctly classifies 738 samples and misclassifies 14 samples, yielding an
overall test accuracy of 98.14%. The per-class results are summarized in Table 8.
Table 7. Per-class classification performance of the proposed Custom Lightweight CNN.
Class Precision Recall F1-Score Support
BB (Brown Blight) 0.95 0.98 0.97 188
GL (Healthy Leaves) 0.99 0.99 0.99 188
RR (Red Rust) 1.00 0.99 1.00 188
RSM (Red Spider Mite) 0.98 0.95 0.97 188
Weighted Average 0.98 0.98 0.98 752
https://doi.org/10.1371/journal.pone.0363128.t007
Fig 5. Confusion matrix of the proposed Custom Lightweight CNN on the tea leaf disease test set.
https://doi.org/10.1371/journal.pone.0363128.g005

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 19 / 33
4.5 ROC curve and AUC analysis
Fig 6 presents the multi-class ROC curves of the proposed Custom Lightweight CNN, computed using a one-vs-rest strat-
egy for each of the four classes. The dashed diagonal line represents the random classifier baseline (AUC = 0.5). All four
class-specific curves rise sharply toward the top-left corner at extremely low false positive rates, indicating strong discrimi-
native capability across all categories.
GL (Healthy Leaves) and RR (Red Rust) attain a perfect AUC of 1.0000, reflecting their visually distinct characteris-
tics and near-zero misclassification rates. BB (Brown Blight) achieves an AUC of 0.9990, while RSM (Red Spider Mite)
achieves 0.9989 — the lowest among the four classes, consistent with RSM recording the highest misclassification count
(9 samples misclassified as BB) in the confusion matrix analysis. The marginal deviations from perfect AUC in BB and
RSM are attributable to the visual similarity between these two classes, where RSM damage and BB lesions share over-
lapping reddish-brown discoloration patterns. Overall, all four classes achieve AUC values at or above 0.9989, confirming
that the proposed model maintains excellent class separability across all tea leaf disease categories.
Table 8. Per-class classification performance derived from the confusion matrix.
Class True Samples Correctly Classified Misclassified Recall (%) Precision (%)
BB (Brown Blight) 188 185 3 (→ RSM) 98.40 94.87
GL (Healthy Leaves) 188 187 1 (→ BB) 99.47 99.47
RR (Red Rust) 188 187 1 (→ GL) 99.47 100.00
RSM (Red Spider Mite) 188 179 9 (→ BB) 95.21 98.35
Total/ Overall 752 738 14 98.14 —
https://doi.org/10.1371/journal.pone.0363128.t008
Fig 6. Multi-class ROC curves of the proposed Custom Lightweight CNN for four-class tea leaf disease classification.
https://doi.org/10.1371/journal.pone.0363128.g006

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 20 / 33
4.6 Train vs. validation learning curves
Fig 7 present the training and validation accuracy and loss curves of the proposed Custom Lightweight CNN over the full
course of training. Together, these two curves serve as a diagnostic tool for assessing the model’s convergence behavior,
stability, and generalization performance.
Accuracy Curves, at epoch 0, the training accuracy begins at approximately 0.60 while the validation accuracy drops
to approximately 0.24, reflecting the typical instability of the initial parameter updates before the model begins to general-
ize. By epoch 2, the validation accuracy recovers rapidly to approximately 0.97, briefly surpassing the training accuracy.
This early crossover is a positive indicator, confirming that the model does not enter an overfit state from the outset.
From epoch 5 onward, both curves converge smoothly and rise together in a stable, monotonic fashion, with the training
accuracy reaching approximately 0.91 and the validation accuracy approximately 0.96 by epoch 5. The training accuracy
stabilizes in the range of 0.97–0.98 from approximately epoch 20, while the validation accuracy stabilizes slightly higher at
0.98–0.99 from approximately epoch 15.
Loss Curves,the training loss begins at approximately 1.01 and the validation loss starts at approximately 1.40 at epoch
0. Both losses decrease steeply in the early epochs: by epoch 3, the validation loss falls to approximately 0.10, while
Fig 7. Training and validation accuracy and loss curves of the proposed Custom Lightweight CNN.
https://doi.org/10.1371/journal.pone.0363128.g007

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 21 / 33
the training loss reaches approximately 0.25 by epoch 5. Minor oscillations in the validation loss are observed between
epochs 5 and 10, reaching as high as 0.13, before stabilizing smoothly from epoch 10 onward. From epoch 15 through
termination, both losses stabilize — the training loss at approximately 0.07–0.10 and the validation loss at approximately
0.03–0.05. Notably, the validation loss remains consistently below the training loss throughout the entire training process,
which is a strong indicator that the model generalizes well to unseen data and exhibits no overfitting.
Convergence and Stopping. Training is automatically halted at approximately epoch 39 by the EarlyStopping callback
(patience = 7), which monitors validation loss and restores the best model weights. The ReduceLROnPlateau callback
(patience = 3, factor = 0.5) further contributes to smooth convergence by progressively reducing the learning rate whenever
improvement plateaus, preventing oscillatory behavior in later epochs. The tight alignment of both accuracy curves and
the consistently lower validation loss relative to training loss from epoch 10 through termination confirm that the proposed
model achieves stable convergence without overfitting, underfitting, or erratic generalization behavior throughout training.
4.7 Grad-CAM faithfulness evaluation: Insertion and deletion curves
To quantitatively assess the faithfulness of the Grad-CAM explanations, insertion and deletion metrics were employed, as
illustrated in Fig 8. These evaluation methods measure how accurately the Grad-CAM heatmaps identify image regions
that are truly important for the model’s prediction. In the insertion experiment, the most salient regions highlighted by
Grad-CAM were progressively introduced into a blurred image, while the predicted probability for the target class was
recorded. As shown by the green curve, the prediction confidence increased sharply from nearly zero to almost one
after only a small fraction of the most important patches were inserted, resulting in a high insertion AUC of 0.972. This
rapid increase indicates that the highlighted regions contain highly discriminative information essential for the model’s
decision-making.
Conversely, the deletion experiment progressively removed the most salient regions from the original image. The red
curve demonstrates a rapid decline in prediction confidence from 1.0 to nearly 0, yielding a low deletion AUC of 0.181.
Fig 8. Grad-CAM faithfulness evaluation of the proposed Custom Lightweight CNN using insertion and deletion curves.
https://doi.org/10.1371/journal.pone.0363128.g008

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 22 / 33
The steep decrease confirms that removing the regions identified by Grad-CAM significantly degrades the model’s confi-
dence, indicating that these regions are indeed critical for accurate classification.
Overall, the combination of a high insertion AUC (0.972) and a low deletion AUC (0.181) demonstrates the strong
faithfulness of the Grad-CAM explanations. The results indicate that the generated saliency maps accurately capture
the image regions that drive the model’s predictions, providing quantitative evidence that the proposed model’s visual
explanations are both reliable and interpretable. These findings complement the qualitative Grad-CAM visualizations and
enhance confidence in the transparency and trustworthiness of the proposed plant disease classification framework.
4.8 Ground truth vs. prediction visualization
Fig 9 presents a sample visualization of the proposed Custom Lightweight CNN’s ground truth labels and corresponding
predictions for 16 representative test images from the BB (Brown Blight) class. For each sample, the ground truth label
(GT), predicted label (Pred), and prediction confidence score (Conf) are displayed, with correct predictions shown in green
and misclassifications in red.
Out of the 16 samples, 15 are correctly classified as BB with confidence scores ranging from 0.64 to 1.00, with the
majority exceeding 0.90 — reflecting the model’s strong recognition of characteristic BB features such as brown discolor-
ation patches and necrotic spots. Four samples achieve a perfect confidence of 1.00. The single lowest confidence among
correct predictions is 0.64, corresponding to a sample with relatively subtle lesion features, yet the model still assigns the
correct class label.
Only one sample is misclassified, predicted as RSM (Red Spider Mite) with a confidence of 0.92. The relatively high
confidence of this misclassification suggests that the sample exhibits strong visual characteristics of RSM — such as dif-
fuse brownish mottling and stippling patterns — that overlap with BB lesion morphology, making it a genuinely ambiguous
case at the BB–RSM decision boundary.
Overall, the visualization confirms that the proposed model performs reliably on the BB class, with the single misclassi-
fication attributable to genuine inter-class visual similarity rather than systematic model failure — reinforcing that the mod-
el’s predictions are both accurate and well-calibrated.
4.9 Grad-CAM explainability analysis
Fig 10 presents Grad-CAM heatmap visualizations for four representative BB (Brown Blight) test samples, each
correctly classified by the proposed Custom Lightweight CNN. For each pair, the left column shows the original
image and the right column shows the Grad-CAM overlay, where warmer colors (green, yellow) indicate regions
of higher influence on the model’s classification decision and cooler colors (blue, purple) indicate regions of lower
importance.
In the first three samples, each leaf displays a single, small, well-defined brownish lesion spot. In all three cases, the
Grad-CAM heatmap produces a tight, localized green/cyan hotspot precisely centered on the lesion, while the surrounding
healthy leaf tissue remains in cool blue and purple tones. This consistent pattern demonstrates that the model correctly
identifies and localizes the individual lesion site as the primary discriminative region for BB classification, even when the
lesion occupies only a small fraction of the total leaf area.
The fourth sample presents two overlapping leaves with multiple reddish-purple lesion spots distributed across their
surfaces. The corresponding Grad-CAM heatmap shows a more diffuse, cool blue-purple activation across both leaves
rather than a single dominant hotspot, reflecting the model’s broader spatial attention in response to the distributed
multi-lesion pattern. This behavior is consistent with the spatially spread nature of the disease presentation in this sample.
To further assess model interpretability, Grad-CAM visualizations were extended to include correctly classified and mis-
classified samples from all four classes, as shown in Fig 11. The figure presents the original images, actual and predicted
labels, confidence scores, and corresponding heatmaps.

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 23 / 33
Overall, the Grad-CAM visualizations demonstrate that the proposed model adapts its attention pattern to the lesion distribution
of each sample — producing precise, localized activation for single-lesion cases and broader distributed activation for multi-lesion
presentations. In all cases, the model’s attention is directed toward the disease-affected regions of the leaf rather than background
or irrelevant areas, confirming the interpretability and reliability of the proposed lightweight architecture for BB disease recognition.
4.10 Ablation study and computational cost analysis
To evaluate the effectiveness of the proposed lightweight CNN, an ablation study was performed by comparing it with nine
widely used CNN architectures. As shown in Fig 12, the proposed model achieved 98.14% accuracy, 98.17% precision,
Fig 9. Ground truth labels, predicted labels, and confidence scores for representative Brown Blight (BB) test samples.
https://doi.org/10.1371/journal.pone.0363128.g009

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 24 / 33
98.14% recall, and 98.14% F1-score, outperforming VGG16, ResNet50, ResNet101, EfficientNetB1, and NASNetMobile,
while remaining competitive with DenseNet121, MobileNetV2, InceptionV3, and Xception. These results demonstrate that
the proposed architecture maintains robust and balanced classification performance despite its compact design. The com-
putational efficiency comparison is presented in Fig 13. The proposed model requires only 0.08 M trainable parameters,
0.31 MB of storage, and 191.29 MFLOPs, with the shortest training time (6820 s) among all evaluated models. Compared
with conventional CNNs, the proposed architecture substantially reduces memory usage and computational complexity
while maintaining high classification accuracy. This favorable balance between predictive performance and computational
cost makes the model well suited for deployment in real-time and resource-constrained environments.
To further investigate the contribution of individual architectural components, a module-level ablation study was con-
ducted by removing depthwise separable convolution (DSC), global average pooling (GAP), and Dropout individually.
As reported in Table 9, removing DSC increased the parameter count and computational cost, while replacing GAP
substantially increased the model size. Removing Dropout improved the observed accuracy to 99.20% without changing
the parameter count or FLOPs. The complete LightTeaNet achieved 98.14% accuracy with only 80,959 parameters and
a model size of 0.31 MB, demonstrating its favorable balance between classification performance and computational
efficiency.
4.11 On-device hardware evaluation
To further evaluate the practical deployment potential of LightTeaNet, on-device inference testing was conducted on a
Xiaomi Redmi Note 7 Pro running Android 10 using the TensorFlow Lite (float32) model. The model achieved a mean
total processing latency of 20.82 ms per image, with a standard deviation of 0.55 ms and a throughput of 48.03 FPS. The
Fig 10. Grad-CAM heatmaps highlighting disease-relevant regions in representative BB test images.
https://doi.org/10.1371/journal.pone.0363128.g010

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 25 / 33
average preprocessing and inference times were 4.26 ms and 16.56 ms, respectively, while peak RAM usage reached
701.56 MB during testing. These results provide practical evidence of LightTeaNet’s feasibility on mobile hardware. How-
ever, validation on dedicated edge-computing devices, such as Raspberry Pi, and under real-world agricultural conditions
remains an important direction for future work.
4.12 Deployment of the proposed tea leaf disease classification system
To demonstrate the practical applicability of the proposed lightweight CNN, a web-based decision support system was
developed using the Streamlit framework. As shown in Fig 14, the application enables users to upload tea leaf images,
automatically performs image preprocessing and validation, and executes real-time disease classification using the
proposed model. The system displays the predicted disease class, confidence score, class probability distribution, and
Fig 11. Grad-CAM visualizations of correctly classified and misclassified tea leaf images, including actual labels, predicted labels, confidence
scores, and attention heatmaps.
https://doi.org/10.1371/journal.pone.0363128.g011

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 26 / 33
uploaded image preview through an intuitive interface. Its lightweight design enables efficient inference with low computa-
tional overhead, making it suitable for practical precision agriculture applications.
4.13 Comparison with Existing State-of-the-Art Methods
As highlighted in Table 10, our proposed method outperforms existing solutions for tea leaf disease detection across
accuracy, model compactness, and computational efficiency. Reference [18] proposed WaveLiteNet, achieving an accu-
racy of 98.70% with 3.16 × 10⁶ parameters and a model size of 35 MB, considerably larger than our model’s 80,959
parameters and 0.31 MB. Reference [17] employed YOLOv8-RCAA, attaining 98.14% mAP with 0.81 M parameters and
6.46 MB model size, while requiring 27.23 M FLOPs. Reference [22] introduced FCHE-YOLO, reaching 98.1% mAP with
1.46 M parameters and 4.2 G FLOPs. Reference [25] utilized WMC-RTDETR, achieving 97.7% mAP50 but requiring
12.82 M parameters, 27.5 MB model size, and 33.9 G FLOPs — substantially higher computational overhead than the
proposed model. Reference [29] applied I-MobileNetV2, obtaining 96.12% accuracy with 1.35 M parameters, whereas
Reference [28] used MobileNetV3-CA, achieving a comparatively lower accuracy of 95.88%. Reference [37] employed
YOLOv8-ASFF, achieving 95.26% mAP with 25.68 M parameters and 74.2 G FLOPs, reflecting a considerably heavier
architecture. Reference [16] combined CNN and MLP, achieving 95.06% accuracy, while Reference [30] used a stan-
dard MobileNetV2, achieving 94.55% accuracy with 2.00 M parameters and a training time of 19 min 1 s. Reference [31]
Fig 12. Performance comparison of the proposed lightweight CNN and benchmark deep learning models.
https://doi.org/10.1371/journal.pone.0363128.g012

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 27 / 33
proposed BHC-YOLOV8, achieving 94.5% mAP0.5 with 12.1 G FLOPs, and Reference [38] introduced TCSA, achieving
94.6% mAP. Reference [26] employed ECA-ResNet50, achieving 93.06% accuracy, while Reference [39] used YOLOv8-
RMDA, achieving 93.04% mAP50. Reference [34] applied an Improved YOLOv7 model, achieving 92.1% mAP with 24.33
M parameters and 37.7 G FLOPs, and Reference [27] combined YOLOv7-tiny with TSCODE, achieving 91.5% mAP0.5.
Reference [35] used EfficientNet, achieving 91% accuracy with a training time of 29 min, while Reference [36] proposed
YOLO-DBD, achieving 86.8% mAP0.5 with 11.6 M parameters and 25.1 G FLOPs. Reference [33] employed YOLOv10s,
achieving 84% mAP50. Reference [23] used a custom deep learning model, reporting the highest comparable accuracy of
99%, though without disclosed parameter, FLOPs, or model size information for direct efficiency comparison. Reference
[20] applied a standard CNN, achieving 96.65% accuracy with a training time of 13 h — considerably longer and less
Fig 13. Computational cost comparison of the proposed lightweight CNN and benchmark models.
https://doi.org/10.1371/journal.pone.0363128.g013
Table 9. Module-level ablation study of LightTeaNet.
Model DSC GAP Dropout Accuracy (%) Parameters FLOPs Model Size (MB)
Full LightTeaNet ✓ ✓ ✓ 98.14 80,959 191,290,328 0.31
Without DSC ✗ ✓ ✓ 99.34 423,748 1,486,132,184 1.62
Without GAP ✓ ✗ ✓ 98.67 6,470,719 204,019,672 24.68
Without Dropout ✓ ✓ ✗ 99.20 80,959 191,290,328 0.31
https://doi.org/10.1371/journal.pone.0363128.t009

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 28 / 33
efficient than the proposed model. Reference [24] utilized an ensemble model, achieving 96% accuracy, and Reference
[15] proposed LTDDN, achieving 98.03% accuracy with 2.998 M parameters and 394.301M FLOPs.
In contrast, the proposed Custom CNN achieves the highest classification accuracy of 98.14% while requiring only
80,959 parameters, a model size of merely 0.31 MB, and 191,290,328 FLOPs figures that are substantially lower than the
majority of the compared methods. Although the proposed model requires a relatively longer training time of 6820.02s,
its markedly smaller footprint and superior accuracy make it highly suitable for deployment on resource-constrained and
edge devices for real-time tea leaf disease diagnosis. Furthermore, the incorporation of Grad-CAM-based explainability
confirms that the proposed model’s predictions are both accurate and interpretable, aligning with the visual explanation
capability offered by several competing methods.
5. Discussion
5.1. Interpretation of Classification Performance
The proposed lightweight CNN achieved high classification performance while maintaining remarkable computational
efficiency, demonstrating that effective architectural optimization can compensate for reduced network complexity. The
integration of depthwise separable convolutions, progressive feature expansion, and Global Average Pooling enables the
model to learn discriminative disease-specific representations while minimizing redundant computations. The consistently
high performance across all disease categories indicates strong feature generalization and stable learning capability. Most
Fig 14. Deployment framework of the proposed lightweight CNN-based tea leaf disease classification system.
https://doi.org/10.1371/journal.pone.0363128.g014

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 29 / 33
misclassifications occurred between visually similar disease classes, particularly those exhibiting overlapping lesion mor-
phology and texture patterns, highlighting the intrinsic difficulty of fine-grained disease recognition rather than limitations of
the proposed model. These findings suggest that the network successfully captures biologically relevant visual character-
istics and provides reliable disease discrimination under controlled imaging conditions.
Table 10. Benchmark comparison of the proposed model against existing tea leaf disease classification methods.
Ref Number of datasets Augmentation
(Yes/no)
Best Model Accuracy Total
Parameter
Model
Size
Total
FLOPS
Training
Time
Grad
Cam/XAl
[18] 3,260 samples Yes WaveLiteNet 98.70% 3.16 × 106 35 MB 0.46 × 109 45 ms
(Inference)
–
[17] 4001 images Yes YOLOv8-RCAA 98.14%
(mAP)
0.81 M 6.46 MB 27.23 M 0.035 s Yes
(Grad-CAM)
[22] 3037 original (5409
augmented)
Yes FCHE-YOLO 98.1%
(mAP)
1.46 M 3.04 MB 4.2 G 47.5 FPS
(Inference)
Yes
(Heatmap)
[25] 160 original (1280
augmented)
Yes WMC-RTDETR 97.7%
(mAP50)
12.82 M 27.5 MB 33.9 G 58.8 FPS Yes
(EigenCAM)
[29] 2544 images Yes I-MobileNetV2 96.12% 1.35 M – 0.21 G – –
[28] 22,380 images Yes MobileNetV3-CA 95.88% 2.7 × 106 – – – Yes
(Heatmap)
[37] 2700 images (3743
labels)
Yes YOLOv8-ASFF 95.26%
(mAP)
25.68 M – 74.2 G 89 FPS
(Inference)
Yes
(Grad-CAM)
[16] 2800 original (5867
augmented)
Yes CNN + MLP 95.06% – – – – –
[30] 5,980 images Yes MobileNetV2 94.55% 2.00 M – – 19 min 1 s –
[31] 4000 images (10000
augmented)
Yes BHC-YOLOV8 94.5%
(mAP0.5)
– – 12.1 G – –
[38] 776 images (3676
augmented)
Yes TCSA 94.6%
(mAP)
– – – – Yes
(Attention)
[26] 885 images Yes ECA-ResNet50 93.06% – – – – –
[39] 4650 images – YOLOv8-RMDA 93.04%
(mAP50)
– – – 132 FPS Yes
(Grad-CAM)
[34] 4280 images Yes Improved
YOLOv7
92.1%
(mAP)
24.33 M 50.11
MB
37.7 G – –
[27] 182 original (3130
augmented)
Yes YOLOv7
tiny + TSCODE
91.5%
(mAP0.5)
– – – – Yes
(Grad-CAM)
[35] 60 images Yes EfficientNet 91% – – – 29 min –
[36] 1108 (8870
augmented)
Yes YOLO-DBD 86.8%
(mAP0.5)
11.6 M – 25.1 G – –
[33] 1000 images Yes YOLOv10s 84%
(mAP50)
– – – – –
[23] 4000 images Yes Custom Deep
Learning Model
99% – – – – –
[20] 3330 images Yes CNN 96.65% – – – 13 h –
[24] 885 images Yes Ensemble Model 96% – – – – Yes
(Grad-CAM)
[15] 10099 images Yes LTDDN 98.03% 2.998 M – 394.301M – YES
Pro-
posed
5000 Yes Custom CNN 98.14% 80959 0.31 191290328 6820.02s Yes
(Grad-CAM)
https://doi.org/10.1371/journal.pone.0363128.t010

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 30 / 33
5.2. Computational efficiency and practical deployment implications
One of the primary contributions of this study is the outstanding balance between classification accuracy and computa-
tional efficiency. Unlike conventional transfer learning models that require millions of parameters, the proposed architec-
ture significantly reduces model complexity while maintaining competitive predictive performance. The compact model
size and low computational cost make it highly suitable for deployment on resource-constrained platforms, including
smartphones, embedded systems, UAVs, and edge-based IoT devices. Such computational efficiency is particularly
important for precision agriculture, where rapid inference, low energy consumption, and offline operation are essential for
real-time field applications. Therefore, the proposed framework provides a practical solution for scalable and cost-effective
tea disease monitoring.
5.3. Model Interpretability and Trustworthiness
Model interpretability is essential for improving user confidence and supporting practical adoption in agricultural appli-
cations. The Grad-CAM visualizations demonstrate that the proposed model consistently focuses on disease-affected
regions rather than irrelevant background information, indicating that its predictions are driven by meaningful pathological
characteristics. Furthermore, the quantitative Insertion and Deletion evaluations confirm the faithfulness of the generated
attention maps, providing objective evidence that the highlighted regions contribute directly to classification decisions. This
combination of qualitative and quantitative explainability enhances the transparency, reliability, and trustworthiness of the
proposed framework, making it more suitable for supporting decision-making in precision agriculture.
5.4. Comparison with state-of-the-art methods
Compared with recently reported lightweight CNNs and transfer learning approaches, the proposed framework achieves
competitive classification performance while requiring substantially fewer computational resources. This demonstrates that
carefully designed domain-specific architectures can effectively preserve feature representation without relying on exces-
sively deep or computationally expensive networks. Rather than maximizing network complexity, the proposed model
emphasizes efficient feature learning and architectural optimization, resulting in an improved balance between predictive
performance and deployment feasibility. These findings further support the growing importance of lightweight deep learn-
ing models for practical agricultural AI applications.
5.5. Limitations and future directions
Despite its promising performance, several limitations should be acknowledged. The model was evaluated using images
collected under controlled conditions and limited to three common tea leaf diseases and one healthy class which may
not fully represent the variability encountered in real agricultural environments. Performance under varying illumination,
complex backgrounds, and disease severity levels should therefore be investigated. Additionally, the current framework
performs image-level classification without lesion localization or severity estimation. The use of isolated leaf images with
uniform black backgrounds limits the assessment of the model’s robustness against field-related challenges, including
background noise, shadows, and overlapping leaves. Due to the limited availability of suitable field-image datasets and
practical constraints, validation under these complex conditions was beyond the scope of the present study. Conse-
quently, the model’s generalization to UAV-based monitoring and field-sensor applications remains to be verified. Future
research should validate the proposed model on larger field-acquired datasets, incorporate additional disease categories,
investigate model compression techniques such as quantization and pruning, and extend the framework toward real-time
detection and mobile edge deployment.
The current framework performs image-level disease classification and does not explicitly localize lesions or quantify
disease severity. A natural extension would incorporate a segmentation or object detection module to localize symptomatic

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 31 / 33
regions, followed by lesion-area estimation or severity grading. Segmentation-based approaches could estimate the pro-
portion of affected leaf area and map it to severity levels, while detection-based architectures could identify multiple symp-
tomatic regions on a single leaf or across overlapping leaves under field conditions. Prior work on tea disease severity
detection [37] indicates that such approaches are feasible. However, they require pixel- or region-level annotations, which
are not available in the TeaLeafNet dataset, and are therefore beyond the scope of the present classification study. The
Grad-CAM maps in this study provide only coarse, class-discriminative localization and should not be interpreted as lesion
segmentation. Integrating the proposed lightweight backbone into a detection or segmentation pipeline, while preserving
its low computational cost, is a direction for future work.
5.6. Implications for sustainable agriculture
The proposed lightweight framework contributes to precision agriculture by enabling rapid, accurate, and computationally
efficient disease diagnosis. Its low memory requirement and high inference efficiency facilitate deployment on afford-
able edge devices, reducing dependence on cloud computing and expensive hardware. Early disease identification can
support timely intervention, reduce unnecessary pesticide application, and improve crop productivity. Consequently, the
proposed model offers a practical and scalable solution for intelligent crop monitoring, promoting sustainable agricultural
management and supporting the broader adoption of AI-driven farming technologies.
5.7 Conclusion
This paper has presented a lightweight, interpretable convolutional neural network for four-class tea leaf disease classi-
fication that achieves exceptional computational efficiency (80,959 parameters; 0.31 MB model size; 191.29M FLOPs)
while maintaining competitive classification accuracy (98.14%). Through systematic comparison with nine transfer learning
baselines and extensive benchmarking against 25 + state-of-the-art methods, we have demonstrated that the proposed
architecture achieves the best trade-off between accuracy and computational complexity among currently reported tea
leaf disease classification frameworks. The incorporation of Grad-CAM visualization with quantitative faithfulness eval-
uation establishes a new standard of model transparency for agricultural AI, ensuring that classification decisions are
grounded in diagnostically relevant leaf regions. The model’s extreme compactness, interpretability, and competitive accu-
racy position it as a practical solution for edge-based tea disease surveillance, addressing critical barriers to the real-world
adoption of AI-driven crop monitoring. Web Application: https://tea-leaf-disease-classifier-main.streamlit.app/
Author contributions
Conceptualization: Md Rifat Sarker, JONAYED AL-FARUK, Joyonta Das, Md. Shahabuddin Saju, Ozifatun Jannat Akhi.
Data curation: Md Rifat Sarker, JONAYED AL-FARUK, Joyonta Das, Md. Shahabuddin Saju.
Formal analysis: Md Rifat Sarker, JONAYED AL-FARUK, Kaniz Fatema, Md. Shahabuddin Saju.
Funding acquisition: Md Rifat Sarker, JONAYED AL-FARUK.
Investigation: Md Rifat Sarker, JONAYED AL-FARUK, Kaniz Fatema.
Methodology: Md Rifat Sarker, JONAYED AL-FARUK, Joyonta Das.
Project administration: Md Rifat Sarker, JONAYED AL-FARUK.
Resources: Md Rifat Sarker, Kaniz Fatema.
Software: Md Rifat Sarker, JONAYED AL-FARUK, Kaniz Fatema, Joyonta Das, Md. Shahabuddin Saju.
Supervision: Ozifatun Jannat Akhi.
Validation: Md Rifat Sarker, JONAYED AL-FARUK, Kaniz Fatema, Joyonta Das, Md. Shahabuddin Saju.

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 32 / 33
Visualization: Md Rifat Sarker, JONAYED AL-FARUK, Kaniz Fatema, Joyonta Das, Md. Shahabuddin Saju.
Writing – original draft: Md Rifat Sarker, JONAYED AL-FARUK, Kaniz Fatema, Joyonta Das, Md. Shahabuddin Saju,
Ozifatun Jannat Akhi.
Writing – review & editing: Md Rifat Sarker, JONAYED AL-FARUK, Kaniz Fatema, Joyonta Das, Md. Shahabuddin Saju,
Ozifatun Jannat Akhi.
References

  1. DevelopmentAid. World’s Largest Tea-Producing Countries. https://www.developmentaid.org/news-stream/post/197164/worlds-largest-tea-produc-
    ing-countries. 2025.
  2. ElectroIQ. Tea Statistics: Global Scenario. https://electroiq.com/stats/tea-statistics/. 2026.
  3. Pandey AK, Sinniah GD, Babu A, Tanti A. How the Global Tea Industry Copes With Fungal Diseases - Challenges and Opportunities. Plant Dis.
    2021;105(7):1868–79. https://doi.org/10.1094/PDIS-09-20-1945-FE PMID: 33734810
  4. Ahmed F, Ahad MT, Emon YR. Machine learning-based tea leaf disease detection: A comprehensive review. 2023. https://doi.org/arXiv:2311.03240
  5. Wu Y, Wan S, Wu L, Zhao Y. DAONet-YOLOv8: An Occlusion-Aware Dual-Attention Network for Tea Leaf Pest and Disease Detection. arXiv pre-
    print. 2025. https://doi.org/arXiv:2511.23222
  6. Swapna C, Sree PC, Vaishnavi G, Battu N. Artificial intelligence for crop health monitoring: A comprehensive review of deep learning. International
    Journal of Research Publication and Reviews. 2025;6(12):1145–54.
  7. Zheng H, Wang H, Dong H, Qian Y. A survey of crop disease recognition methods based on spectral and RGB images. J Imaging. 2026;12(2):66.
    https://doi.org/10.3390/jimaging12020066 PMID: 41745431
  8. Pacal I, Kunduracioglu I, Alma MH, Deveci M, Kadry S, Nedoma J, et al. A systematic review of deep learning techniques for plant diseases. Artif
    Intell Rev. 2024;57(11). https://doi.org/10.1007/s10462-024-10944-7
  9. Bagga M, Goyal S. Image‐based detection and classification of plant diseases using deep learning: State‐of‐the‐art review. Urban Agric & Regnl
    Food Sys. 2024;9(1). https://doi.org/10.1002/uar2.20053
  10. Weloday F, Su J. LWMSCNN-SE: A Lightweight Multi-Scale Network for Efficient Maize Disease Classification on Edge Devices. arXiv preprint
    arXiv:2601.07957. 2026.
  11. Murshed MGS, Murphy C, Hou D, Khan N, Ananthanarayanan G, Hussain F. Machine learning at the network edge: A survey. ACM Comput Surv.
    2021;54(8):1–37. https://doi.org/10.1145/3469029
  12. Pal C, Karmakar S, Mukherjee I, Chakrabarti PP. A lightweight and explainable CNN model for empowering plant disease diagnosis. Sci Rep.
    2025;15(1):30720. https://doi.org/10.1038/s41598-025-94083-1 PMID: 40841589
  13. Nyawose T, Maswanganyi RC, Khumalo P. A Review on the Detection of Plant Disease Using Machine Learning and Deep Learning Approaches. J
    Imaging. 2025;11(10):326. https://doi.org/10.3390/jimaging11100326 PMID: 41150002
  14. Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based
    Localization. In: 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 618–26. https://doi.org/10.1109/iccv.2017.74
  15. Liang J, Liang R, Wang D. A novel lightweight model for tea disease classification based on feature reuse and channel focus attention mechanism.
    Engineering Sci Technol Int J. 2025;61:101940. https://doi.org/10.1016/j.jestch.2024.101940
  16. Balasundaram A, Sundaresan P, Bhavsar A, Mattu M, Kavitha MS, Shaik A. Tea leaf disease detection using segment anything model and deep
    convolutional neural networks. Results in Engineering. 2025;25:103784. https://doi.org/10.1016/j.rineng.2024.103784
  17. Wang J, Li M, Han C, Guo X. YOLOv8-RCAA: A Lightweight and High-Performance Network for Tea Leaf Disease Detection. Agriculture.
    2024;14(8):1240. https://doi.org/10.3390/agriculture14081240
  18. Yang J, Xu G, Yang M, Lin Z. Lightweight wavelet-CNN tea leaf disease detection. PLoS One. 2025;20(5):e0323322. https://doi.org/10.1371/jour-
    nal.pone.0323322 PMID: 40424373
  19. Heng Q, Yu S, Zhang Y. A new AI-based approach for automatic identification of tea leaf disease using deep neural network based on hybrid pool-
    ing. Heliyon. 2024;10(5):e26465. https://doi.org/10.1016/j.heliyon.2024.e26465 PMID: 38434404
  20. Rahman H, Ahmad I, Jon PH, Salam A, Rabbi MF. Automated detection of selected tea leaf diseases in Bangladesh with convolutional neural
    network. Sci Rep. 2024;14(1):14097. https://doi.org/10.1038/s41598-024-62058-3 PMID: 38890367
  21. Li J, Liao C. Tea Disease Recognition Based on Image Segmentation and Data Augmentation. IEEE Access. 2025;13:19664–77. https://doi.
    org/10.1109/access.2025.3534024
  22. Zhang Y, Wu G, Shen J. Precise tea leaf disease detection using UAV low-altitude remote sensing and optimized YOLO11 model. PLoS One.
    2026;21(2):e0342545. https://doi.org/10.1371/journal.pone.0342545 PMID: 41706797
  23. Rahat IS, Ghosh H, Dara S, Kant S. Towards precision agriculture tea leaf disease detection using CNNs and image processing. Sci Rep.
    2025;15(1):17571. https://doi.org/10.1038/s41598-025-02378-0 PMID: 40399405

PLOS One | https://doi.org/10.1371/journal.pone.0363128 November 18, 2026 33 / 33
24. Ozturk O, Sarica B, Seker D. Interpretable and Robust Ensemble Deep Learning Framework for Tea Leaf Disease Classification. Horticulturae.
2025;11(4):437. https://doi.org/10.3390/horticulturae11040437
25. Zhang Y, Song J, Yu X, Ji X. WMC-RTDETR: a lightweight tea disease detection model. Front Plant Sci. 2025;16:1574920. https://doi.org/10.3389/
fpls.2025.1574920 PMID: 40395277
26. Li L, Zhao Y. Tea disease identification based on ECA attention mechanism ResNet50 network. Frontiers in Plant Science. 2025;16:1489655.
27. Yao X, Lin H, Bai D, Zhou H. A Small Target Tea Leaf Disease Detection Model Combined with Transfer Learning. Forests. 2024;15(4):591. https://
doi.org/10.3390/f15040591
28. Li Y, Lu Y, Liu H, Bai J, Yang C, Yuan H, et al. Tea leaf disease and insect identification based on improved MobileNetV3. Front Plant Sci.
2024;15:1459292. https://doi.org/10.3389/fpls.2024.1459292 PMID: 39399539
29. Li Z, Li Y, Yan C, Yan P, Li X, Yu M, et al. Enhancing Tea Leaf Disease Identification with Lightweight MobileNetV2. CMC. 2024;80(1):679–94.
https://doi.org/10.32604/cmc.2024.051526
30. Hairah U, Septiarini A, Puspitasari N, Tejawati A, Hamdani H, Eka Priyatna S. Classification of tea leaf disease using convolutional neural network
approach. IJECE. 2024;14(3):3287. https://doi.org/10.11591/ijece.v14i3.pp3287-3294
31. Zhan B, Xiong X, Li X, Luo W. BHC-YOLOV8: improved YOLOv8-based BHC target detection model for tea leaf disease and defect in real-world
scenarios. Front Plant Sci. 2024;15:1492504. https://doi.org/10.3389/fpls.2024.1492504 PMID: 39687320
32. Chen Z, Zhou H, Lin H, Bai D. TeaViTNet: Tea Disease and Pest Detection Model Based on Fused Multiscale Attention. Agronomy. 2024;14(3):633.
https://doi.org/10.3390/agronomy14030633
33. Deng X, Chen T, Photong C. Automated identification of tea leaf diseases and pests using deep learning methods. Operational Research in Eng
Sciences: Theory and Applications. 2024;7(2).
34. Xia Y, Yuan W, Zhang S, Wang Q, Liu X, Wang H, et al. Classification and identification of tea diseases based on improved YOLOv7 model of
MobileNeXt. Sci Rep. 2024;14(1):11799. https://doi.org/10.1038/s41598-024-62451-y PMID: 38782981
35. Kumar, G, & Gupta, A. (2024). An integrated tea leaf diseases identification and retrieval model using machine learning and deep learning
approach.
36. Leng X, Chen J, Huang J, Zhang L, Li Z. An Improved YOLOv8-Based Method for Real-Time Detection of Harmful Tea Leaves in Complex Back-
grounds. Phyton. 2024;93(11):2963–81. https://doi.org/10.32604/phyton.2024.057166
37. Ye R, Shao G, Yang Z, Sun Y, Gao Q, Li T. Detection Model of Tea Disease Severity under Low Light Intensity Based on YOLOv8 and Enlighten-
GAN. Plants (Basel). 2024;13(10):1377. https://doi.org/10.3390/plants13101377 PMID: 38794447
38. Sun Y, Jiang M, Guo H, Zhang L, Yao J, Wu F, et al. Multiscale Tea Disease Detection with Channel–Spatial Attention. Sustainability.
2024;16(16):6859. https://doi.org/10.3390/su16166859
39. Ye R, Shao G, He Y, Gao Q, Li T. YOLOv8-RMDA: Lightweight YOLOv8 network for early detection of small target diseases in tea. Sensors
(Basel). 2024;24(9):2896. https://doi.org/10.3390/s24092896 PMID: 38733002
40. Zhao T, Tong W, Wang Z, Fu C, Gao C. UADC: Uncertainty-aware diverse co-training for semi-supervised medical image segmentation. Expert
Systems. 2026;43(7):e70298. https://doi.org/10.1111/exsy.70298
41. Zhao T, Tong W, Wang Z, Ye C, Gao C, Fu C. RGD ‐ UNet: Efficient decoupled fully connection attention for pixel‐level segmentation of signet‐ring
cells. Expert Systems. 2026;43(9). https://doi.org/10.1111/exsy.70374
42. Dibru HS. TeaLeafNet: Tea Leaf Disease Dataset. https://www.kaggle.com/datasets/harjindersinghdibru/tealeafnet. Accessed 2026 June.

论文原文与出版信息

PLoS ONE
查看论文原文

继续探索