Click through the PLOS taxonomy to find articles in your field.
For more information about PLOS Subject Areas, click here .
Total Mendeley and Citeulike bookmarks.
Paper's citation count computed by Dimensions.
Sum of Facebook, Twitter, Reddit and Wikipedia activity.
Roles Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Validation, Visualization, Writing – original draft, Writing – review & editing
Affiliation Department of Computer Science and Engineering, Mymensingh Engineering College, Mymensingh, Bangladesh
Roles Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Software, Validation, Visualization, Writing – original draft, Writing – review & editing
* E-mail: jonayedalfaruk211282@gmail.com
https://orcid.org/0009-0004-5291-0855
Roles Formal analysis, Investigation, Resources, Software, Validation, Visualization, Writing – original draft, Writing – review & editing
Roles Conceptualization, Data curation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing
Affiliation Department of Electrical and Electronic Engineering, Mymensingh Engineering College, Mymensingh, Bangladesh
Roles Conceptualization, Data curation, Formal analysis, Software, Validation, Visualization, Writing – original draft, Writing – review & editing
Affiliation Department of Civil Engineering, Mymensingh Engineering College, Mymensingh, Bangladesh
Roles Conceptualization, Supervision, Writing – original draft, Writing – review & editing
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 detection systems remain impractical for real-world deployment due to prohibitive computational 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—representing 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, approximately 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. Comparative 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. 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.
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: Data is available at kaggle; https://doi.org/10.34740/kaggle/dsv/11770641 .
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
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 compared 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 dominant 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 performance, most standard CNN architectures are computationally expensive. State-of-the-art architectures such as VGG, ResNet, DenseNet, and EfficientNet achieve high accuracy by leveraging deep and complex network designs; however, these models typically involve millions of parameters 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 overhead, 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 classification. 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 computational 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 Deletion faithfulness evaluation.
The main contributions of this paper are as follows:
To begin, Liang et al. (2025) [ 15 ] proposed LTDDN, a lightweight deep learning framework for tea disease classification 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% accuracy 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 categories 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 collected 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 framework that combines a Hybrid Pooling CNN for feature extraction with a Weighted Random Forest (WRF) classifier optimized 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% precision. 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 zooming, shearing, and flipping. Despite its high performance, the study was limited to four classes and relied on a custom CNN architecture without exploring advanced transfer-learning models.
Meanwhile, addressing data scarcity through generative augmentation, Li and Liao (2025) [ 21 ] proposed a tea disease 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 dataset, 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 interpretability. 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 computational 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 blurring. 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, 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 learning 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 similar 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 enhancement, 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. Experimental 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 representation 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 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 classification. 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 several 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, channel 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 backgrounds.Ye et al. (2024) [ 39 ] introduced YOLOv8-RMDA, a lightweight tea disease detection model that integrates RFCBAM, 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 computational 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 learning strategies for complex visual recognition tasks.
This study proposes a lightweight and computationally efficient Convolutional Neural Network (CNN) framework for automated tea leaf disease classification. The overall workflow consists of seven major stages: data collection, data preprocessing, 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 .
https://doi.org/10.1371/journal.pone.0363128.g001
https://doi.org/10.1371/journal.pone.0363128.g002
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, 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 convolutions. 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 optimizer and categorical cross-entropy loss function. The trained model is evaluated using multiple performance metrics, including accuracy, precision, recall, F1-score, confusion matrix, and ROC analysis. Finally, Grad-CAM visualization is employed to interpret the model’s predictions by highlighting the most discriminative disease regions within the leaf images.
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 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 chromatic 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 contains an isolated tea leaf positioned against a uniform black background, allowing classification models to focus primarily on disease-related visual features.
To address the class imbalance present in the original dataset, a balanced subset containing 5,000 images was utilized 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 .
https://doi.org/10.1371/journal.pone.0363128.t001
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 acquisition, 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 balanced 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.
https://doi.org/10.1371/journal.pone.0363128.t002
The dataset comprising 5,000 images was divided into training, validation, and testing subsets to facilitate model development 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 balanced 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 .
https://doi.org/10.1371/journal.pone.0363128.t003
https://doi.org/10.1371/journal.pone.0363128.g003
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.
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 improving the model’s generalization performance on unseen samples.
Representative examples of the augmented tea leaf images are presented in Fig 4 , illustrating the diversity introduced through the augmentation pipeline while preserving the underlying disease characteristics essential for accurate classification.
https://doi.org/10.1371/journal.pone.0363128.g004
Pixel intensity values were normalized to the range [0,1] using:
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.
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 excessive computational overhead.
The proposed network consists of four sequential convolutional blocks. Each block contains:
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 computational burden.
The operation of each feature extraction block can be represented as:
where (X i ) denotes the input feature map and (F i ) represents the corresponding output feature map.
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 deactivating 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:
where (z i ) denotes the output score associated with class (i).
The predicted disease class is determined by selecting the category with the highest posterior probability.
The proposed CNN was trained using the Adam optimization algorithm owing to its adaptive learning capability and efficient 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:
where (C) represents the number of disease classes, (y i ) denotes the true label, and (ŷ i ) 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.
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 BatchNormalization 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, GlobalAveragePooling2D 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 .
https://doi.org/10.1371/journal.pone.0363128.t004
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 ]:
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 negative. 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 number of trainable parameters (Params) and approximate model size in megabytes (MB). The parameter count of a convolutional layer is formulated as [ 15 ]:
where K h and K w represent the height and width of the convolution kernel, and C in and C out 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.
The adoption of SeparableConv2D as the core building block provides the primary computational advantage in the proposed model. The floating-point operations (FLOPs) of a convolutional layer are formulated as [ 15 ]:
Where H and W represent the height and width of the input feature map, K h and K w represent the convolution kernel dimensions, and C in and C out represent the number of input and output channels, respectively.
In standard convolution, a layer with input channels C in , output channels C out , and kernel size k × k requires k 2 × C in ×C out multiplications per spatial location. In depthwise separable convolution, this cost is reduced to k 2 × C in +C in ×C out , representing a reduction factor of approximately 1/C out +1/k 2 . 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.
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.
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 features. This analysis improves the transparency, reliability, and practical applicability of the proposed framework for tea leaf disease diagnosis.
The proposed framework introduces a lightweight and interpretable CNN architecture specifically designed for tea leaf disease 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 interpretability, 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.
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 categories: 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 ReduceLROnPlateau 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, NASNetMobile, 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.
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 Adam optimizer (learning rate of 0.0001) and categorical cross-entropy loss. The evaluation considers accuracy, precision, 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 MobileNetV2 (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 classification performance.
https://doi.org/10.1371/journal.pone.0363128.t005
The pretrained backbones expect different input conventions: raw [0, 255] values with internal normalization (EfficientNet), mean-subtracted BGR input (ResNet, VGG16), or scaling to [0, 1] or [− 1 ] with normalization (DenseNet121, MobileNetV2, 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 LightTeaNet with MobileNetV2, DenseNet121, and EfficientNetB1, including inference latency. DenseNet121 achieves the highest 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 parameters and computational operations. Although inference latency is unavailable for LightTeaNet, its compact architecture indicates potential for deployment on resource-constrained agricultural devices.
https://doi.org/10.1371/journal.pone.0363128.t006
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.
https://doi.org/10.1371/journal.pone.0363128.t007
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).
https://doi.org/10.1371/journal.pone.0363128.g005
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 .
https://doi.org/10.1371/journal.pone.0363128.t008
Fig 6 presents the multi-class ROC curves of the proposed Custom Lightweight CNN, computed using a one-vs-rest strategy 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 discriminative capability across all categories.
https://doi.org/10.1371/journal.pone.0363128.g006
GL (Healthy Leaves) and RR (Red Rust) attain a perfect AUC of 1.0000, reflecting their visually distinct characteristics 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 overlapping 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.
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.
https://doi.org/10.1371/journal.pone.0363128.g007
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 generalize. 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 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.
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.
https://doi.org/10.1371/journal.pone.0363128.g008
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. The steep decrease confirms that removing the regions identified by Grad-CAM significantly degrades the model’s confidence, 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.
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.
https://doi.org/10.1371/journal.pone.0363128.g009
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 discoloration 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 diffuse 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 misclassification attributable to genuine inter-class visual similarity rather than systematic model failure — reinforcing that the model’s predictions are both accurate and well-calibrated.
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.
https://doi.org/10.1371/journal.pone.0363128.g010
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 misclassified 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.
https://doi.org/10.1371/journal.pone.0363128.g011
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.
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, 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 computational 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.
https://doi.org/10.1371/journal.pone.0363128.g012
https://doi.org/10.1371/journal.pone.0363128.g013
To further investigate the contribution of individual architectural components, a module-level ablation study was conducted 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.
https://doi.org/10.1371/journal.pone.0363128.t009
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 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. However, validation on dedicated edge-computing devices, such as Raspberry Pi, and under real-world agricultural conditions remains an important direction for future work.
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 uploaded image preview through an intuitive interface. Its lightweight design enables efficient inference with low computational overhead, making it suitable for practical precision agriculture applications.
https://doi.org/10.1371/journal.pone.0363128.g014
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 accuracy 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 standard MobileNetV2, achieving 94.55% accuracy with 2.00 M parameters and a training time of 19 min 1 s. Reference [ 31 ] 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 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.
https://doi.org/10.1371/journal.pone.0363128.t010
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.
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 misclassifications occurred between visually similar disease classes, particularly those exhibiting overlapping lesion morphology 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 characteristics and provides reliable disease discrimination under controlled imaging conditions.
One of the primary contributions of this study is the outstanding balance between classification accuracy and computational efficiency. Unlike conventional transfer learning models that require millions of parameters, the proposed architecture 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.
Model interpretability is essential for improving user confidence and supporting practical adoption in agricultural applications. 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.
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 excessively 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 learning models for practical agricultural AI applications.
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. Consequently, 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 regions, followed by lesion-area estimation or severity grading. Segmentation-based approaches could estimate the proportion of affected leaf area and map it to severity levels, while detection-based architectures could identify multiple symptomatic 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.
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 affordable 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.
This paper has presented a lightweight, interpretable convolutional neural network for four-class tea leaf disease classification 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 evaluation 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 accuracy 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/
Is the Subject Area "Tea" applicable to this article? Yes No
Is the Subject Area "Imaging techniques" applicable to this article? Yes No
Is the Subject Area "Convolution" applicable to this article? Yes No
Is the Subject Area "Deep learning" applicable to this article? Yes No
Is the Subject Area "Plant pathology" applicable to this article? Yes No
Is the Subject Area "Leaves" applicable to this article? Yes No
Is the Subject Area "Mites" applicable to this article? Yes No
Is the Subject Area "Preprocessing" applicable to this article? Yes No
PLOS is a nonprofit 501(c)(3) corporation, #C2354500, based in California, US
Loading comments…