A Lightweight Deep Learning Model for Tea Leaf Disease Identification

Abstract

Tea is a globally important economic crop, and the ability to quickly and accurately identify tea leaf diseases can significantly improve both the yield and quality of tea production. With advances in deep learning, many recent studies have demonstrated that convolutional neural networks are both feasible and effective for identifying tea leaf diseases. In this paper, we propose a modified EfficientNetB0 lightweight convolutional neural network, enhanced with the ECA module, to reliably identify various tea leaf diseases. We used two tea leaf disease datasets from the Kaggle platform: the Tea_Leaf_Disease dataset, which contains six categories, and the teaLeafBD dataset, which includes seven categories. Experimental results show that our method substantially reduces computational costs, the number of parameters, and overall model size. Additionally, it achieves accuracies of 99.49% and 90.73% on these widely used datasets, making it highly suitable for practical deployment on resource-constrained edge devices.

1. Introduction

Tea leaves are one of the world’s most important cash crops, with an annual production of millions of tons [1]. After processing, tea buds become one of the world’s most popular beverages, with global consumption reaching approximately 20 billion cups per day [2]. This massive demand makes tea a key agricultural export for many countries and the primary source of livelihood for millions of tea farmers. However, tea diseases can cause significant declines in both quality and yield, leading to severe economic losses for farmers [3]. The types of tea leaf diseases are diverse, with seven common diseases being Algal Spot, Brown Blight, Gray Blight, Helopeltis, Red Spot, Red Spider, Green Mirid Bug, etc. These diseases are especially prevalent in warm, humid climates and are primarily caused by insect infestations and fungal infections. When disease outbreaks are severe, they often result in extensive leaf drop, which reduces harvest yields. The early identification of tea diseases can help mitigate these losses, thereby improving the overall production and quality of tea. For these reasons, research into tea leaf diseases is vital for the sustainable development and economic stability of the tea industry.

With the advancement of deep learning technology, convolutional neural networks (CNNs) have gradually replaced traditional manual methods in plant leaf disease identification. Manual identification requires expert evaluation, which is often time-consuming and costly. In contrast, CNN-based automated identification offers the advantages of speed and high accuracy, demonstrating remarkable success in the field of plant leaf disease detection [4,5,6,7]. Currently, CNNs are among the most widely used and effective techniques for identifying plant leaf diseases. By analyzing correlations between image pixels, they can detect features such as spots and color changes, enabling accurate disease classification. In recent years, CNNs have been extensively applied to tea leaf disease identification, significantly reducing labor costs and minimizing misdiagnoses, thereby helping to protect the growth and yield of tea leaves. Although CNN models have achieved significant success in plant disease recognition, their practical deployment still faces several challenges. These include (1) reducing computational and resource demands, (2) ensuring effective deployment in real-world agricultural environments, (3) lowering model complexity, and (4) accurately distinguishing visually similar diseases. To address these issues, various enhancement approaches have been proposed in the related research.

Singh et al. [8] proposed a tea leaf disease identification model based on an improved deep convolutional neural network designed to accurately and rapidly detect tea leaf diseases, facilitating their prevention and management. The model architecture consists of three convolutional layers, a max pooling layer, a flattening layer, and three fully connected layers. Data augmentation techniques were applied during training to enhance model performance. Experimental results show that for identifying eight types of single-background tea leaf conditions, an average recognition accuracy of 73% is achieved.

Wu et al. [9] addressed the challenge of rapidly identifying tea leaf diseases by proposing a model called LBPAttNet. This model combines a convolutional neural network for extracting deep features with local binary patterns to capture local texture details for accurate identification. Using ResNet18 as the backbone, the model incorporates a lightweight Coordinate Attention (CA) mechanism to focus more precisely on disease-affected regions of the tea leaves. To address class imbalance in the tea_sickness_dataset [10], a focal loss function is employed during training. Experimental results demonstrate that the proposed approach achieves average recognition accuracies of 92.78% in the tea_sickness_dataset and 98.13% in the Tea_Leaf_Disease dataset, respectively.

Bhuyan et al. [11] proposed a lightweight model called Res4net-CBAM for automated detection of tea leaf diseases. This model integrates the Res4net architecture, which consists of four residual layers, with the Convolutional Block Attention Module (CBAM) [12]. The dataset images used in their experiments were captured directly in a local tea garden environment and combined with three categories from the tea_sickness_dataset: Brown Blight, Red Spot, and healthy. The experimental results show that, compared with classic convolutional neural networks such as AlexNet, VGG16, ResNet50, DenseNet121, and InceptionV3, the approach achieves a best recognition accuracy of 98.27%.

Latha et al. [13] applied a CNN model to enhance the accuracy of tea leaf disease identification. Their CNN architecture consists of one input layer, four convolutional layers, and two fully connected layers. Some image classes in their experiments were captured directly from the field using a camera, and the dataset was further expanded by incorporating CIFAR-10 along with data augmentation techniques. Using eight types of tea leaf disease images, they further improved the model’s performance by tuning the number of layers, learning rate, and optimizers. As a result, the model achieved an average identification accuracy of 94.45%.

Ozturk et al. [14] integrated deep learning architectures—including ResNet50, MobileNet, EfficientNetB0, and DenseNet121—to identify tea leaf diseases, effectively addressing the challenges posed by complex backgrounds and similarities among different disease types. They applied data augmentation techniques to increase data diversity and trained their models using the tea_sickness_dataset. Using 5-fold cross-validation for evaluation, their approach achieved a recognition accuracy of 96%.

Li et al. [15] addressed the need for lightweight yet highly accurate recognition models suitable for practical tea production. They proposed an improved MobileNetV3 model enhanced with a CA module [16] and employed transfer learning techniques. They applied data augmentation techniques to assist training on a dataset of 17 types of tea leaf diseases and pests collected from real farmland. When compared with models such as ResNet18, AlexNet, VGG16, SqueezeNet, and ShuffleNetV2, their method achieved a best recognition accuracy of 95.88%, and it has already been deployed in actual tea plantations.

Heng et al. [17] proposed a CNN that employs hybrid pooling to automatically and accurately identify tea leaf diseases by effectively extracting their features. In this approach, the pooling layer dynamically combines max pooling and average pooling in a weighted random manner. For disease classification, a weighted random forest model is used, with its weights optimized through the cuckoo search optimization algorithm. Experimental results on the tea_sickness_dataset demonstrate an average identification accuracy of 92.47%.

Chen et al. [18] proposed a lightweight model, TeaViTNet, for detecting tea leaf diseases and pests, addressing the challenges of detecting lesions and pest manifestations at different scales. The model adopts the lightweight transformer architecture MobileViT as the backbone network to capture global features. At the neck, a multi-scale attention mechanism (EMA-PANet) is integrated to enhance feature learning across scales, while an RFBNet module is incorporated to improve the detection of subtle feature differences. To further improve efficiency, an ODCSP convolutional layer is introduced into PANet. Experimental results demonstrated that, on images collected from tea plantations, the model achieved a mean accuracy of 89.1%.

Lin et al. [19] introduced an enhanced real-time detection transformer (RT-DETR) model to achieve automated and efficient tea leaf disease detection. The method incorporates partial convolution (PConv) and a hierarchical design from Faster-LTNet to reduce computational costs, while a CG attention module is integrated to mitigate redundant computations in transformer multi-head self-attention (MHSA) and to strengthen feature representation. Additionally, an RMT spatial prior module is employed to improve the efficiency of both global and local feature representation. Experimental results showed that, on images collected from actual tea plantations, the proposed model achieved a final detection accuracy of 89.2%.

Zhang et al. [20] proposed a lightweight model named WMC-RTDETR to address reduced detection accuracy of tea leaf diseases and pests in complex backgrounds. The model combines the RT-DETR architecture with wavelet transform convolution and introduces a multi-scale multi-head self-attention module to enhance the detection of multi-scale features. Furthermore, a spatial feature reconstruction feature pyramid network was designed to improve adaptability under complex environmental conditions. Experimental results demonstrated that, on natural tea plantation images, the model achieved an mAP50 of 97.7% and an mAP95 of 83.3%.

Kanagarajan et al. [21] developed the AIM-Net model to reduce dependence on large amounts of labeled data, focusing on classifying tea Red Spider mite disease into mild, moderate, and severe levels. The model employs self-supervised learning with swapping assignments between views (SwAV) for pretraining and fine-tuning on a subset, and performance was assessed using 5-fold cross-validation. Experimental results showed that, on images collected from actual farmland, the model achieved a final accuracy of 98.7% while reducing the required labeled data by 62%. Moreover, AIM-Net exhibits a low computational cost and can be deployed on edge devices, highlighting its potential for practical applications in real-world tea plantation disease monitoring.

Most current studies on tea leaf disease identification employ CNNs, reflecting the gradual replacement of traditional manual methods by deep learning techniques in plant disease detection. However, deep neural networks typically require substantial computational power and resources, which limits their practicality for deployment on resource-constrained devices commonly used in real-world agricultural settings. To address this challenge, this paper proposes a model that combines an improved lightweight convolutional neural network—EfficientNetB0 [22]—with an efficient channel attention (ECA) [23] mechanism. We selected two publicly available tea leaf disease image datasets for our experiments. The Tea_Leaf_Disease dataset includes five common tea leaf disease categories plus a healthy leaf category: Algal Spot, Brown Blight, Gray Blight, Helopeltis (tea mosquito bug), and Red Spot. The teaLeafBD dataset contains six common tea leaf disease categories plus a healthy leaf category: Tea Algal Leaf Spot, Brown Blight, Gray Blight, Helopeltis (tea mosquito bug), Red Spider, Green Mirid Bug, and healthy leaf. Our approach aims to accurately identify tea leaf diseases while minimizing computational resource requirements and reducing model complexity. This solution enables efficient and precise disease identification on limited hardware, offering significant potential for advancing smart and precision agriculture. The main contributions of this paper are as follows:

• An improved EfficientNetB0 architecture is proposed, which reduces the number of MBConv modules in Stage 7 and removes the MBConv modules in Stage 8, effectively decreasing model complexity and computational cost while maintaining high classification accuracy.
• The squeeze-and-excitation (SE) attention mechanism in the MBConv modules is replaced with an efficient channel attention (ECA) mechanism, further reducing the number of parameters and improving inference efficiency.
• Extensive experiments conducted on two publicly available tea leaf disease datasets achieved state-of-the-art accuracies of 99.49% and 90.73%, respectively, outperforming other lightweight deep learning models.

2. The Proposed Approach

2.1. Lightweight CNN Model for Tea Leaf Disease Identification

This paper adopts lightweight deep learning techniques to address the problem of tea leaf disease recognition. The overall workflow is illustrated in Figure 1. The tea leaf image dataset was first standardized by resizing all images to a uniform size. The dataset was then split into two parts: one for training the model (training phase) and the other for evaluating its performance (testing phase). During the training phase, we employed 10-fold cross-validation on the training data to balance bias and variance and to reduce the risk of overfitting. Various data augmentation techniques were also employed to increase the diversity of the training samples, improving the model’s adaptability. In the subsequent testing phase, images from the testing set were fed into the trained lightweight convolutional neural network to identify tea leaf diseases. The recognition results were used to evaluate the model’s performance and validate the reliability of the disease identification.

2.2. EfficientNetB0 Architecture

Tea leaf diseases often display similar lesion characteristics in images, requiring the identification model to possess strong discriminative capability. Compared to other lightweight models like MobileNet and ShuffleNet, this paper adopts the EfficientNet architecture to identify tea leaf diseases, aiming to achieve higher accuracy alongside a smaller model size. EfficientNet’s core design philosophy leverages neural architecture search (NAS) to automatically find the optimal balance between network depth, width, and resolution, thereby maximizing accuracy while maintaining computational efficiency [24].

We adopt EfficientNetB0 as the deep learning model for this paper; it is the smallest version in the EfficientNet series. EfficientNetB0 mainly consists of multiple mobile inverted bottleneck convolution (MBConv) modules. As the input image passes through successive MBConv layers, its spatial dimensions are reduced while the number of feature channels increases, allowing the network to extract progressively deeper and more complex features. The architecture of EfficientNetB0 is illustrated in Figure 2.

The MBConv module has two variants based on different expansion factors: MBConv1 and MBConv6. MBConv1 uses the original expansion coefficient of 1, while MBConv6 expands the input feature channels by a factor of 6. The architecture of MBConv1, shown in Figure 3a, begins by extracting features using depthwise separable convolutions, which efficiently capture spatial information with fewer parameters. These features are then enhanced by a squeeze-and-excitation (SE) attention module [25] to emphasize the most informative channels. Next, a 1 × 1 convolution compresses the channels to fuse information across feature maps. To reduce overfitting, a dropout layer is applied. Additionally, residual connections merge input information into the main path, improving feature representation and stabilizing training by mitigating the vanishing gradient problem. MBConv6 follows a similar design but starts by expanding the input channels six-fold through an initial 1 × 1 convolution, as depicted in Figure 3b. This expansion allows the module to extract richer and deeper features before applying depthwise convolution, SE attention, and channel compression. These MBConv modules provide an efficient and effective feature extraction mechanism within EfficientNetB0, with the integrated SE module enhancing feature quality and the overall model’s ability to handle complex visual patterns typical in tea leaf disease images.

However, to meet the strict computational resource and processing speed constraints of edge devices such as smartphones or drones, this paper proposes a lightweight adaptation of the original EfficientNetB0 architecture. This tailored modification aims to retain effective feature extraction capabilities while reducing model complexity and resource demands for practical deployment in resource-constrained environments.

2.3. SE Attention Module

EfficientNetB0 incorporates a squeeze-and-excitation (SE) attention module within each MBConv block. This module autonomously learns the importance of each channel by modeling inter-channel correlations, which helps the model focus on crucial features while suppressing redundant or less important ones. As illustrated in Figure 4, the SE module consists of global average pooling (GAP) followed by two fully connected layers. These layers maintain consistent input and output dimensions and reinforce the significance of inter-channel dependencies. This design enables the model to effectively attend to key features, making it well-suited for tasks like tea leaf disease identification that require precise differentiation of subtle disease variations.

The SE attention module employs two fully connected layers to capture the relationships between channel weights. This approach enhances feature representation by first reducing the dimensionality of the channel descriptor vectors to compress the features. These compressed features then pass through a ReLU activation function before the second fully connected layer restores the dimensionality to its original size. While this two-layer structure strengthens channel feature representation, it also increases the number of parameters and floating-point operations (FLOPs). As a result, it imposes a significant computational cost, posing challenges for deployment on devices with limited resources.

To overcome the increased computational burden mentioned above, this paper replaces the SE module in the MBConv architecture with an ECA module. This avoids the dimension reduction and dimension increase design of the fully connected layer in the SE module, thereby effectively reducing the number of parameters and improving inference efficiency.

2.4. Modified EfficientNetB0 Model Architecture

EfficientNetB0 is recognized for its lightweight design and excellent performance, making it a popular choice in image processing. Its architecture offers high scalability by stacking multiple layers, allowing it to efficiently balance computational resources and model accuracy depending on the specific needs of various recognition tasks. As a result, EfficientNetB0 consistently delivers strong results in deep learning applications and stands out as a promising option among lightweight models.

To enhance deployment efficiency of tea disease identification models in real-world scenarios and ensure their viability on resource-constrained edge devices, this paper introduces a lightweight modification of the EfficientNetB0 architecture, referred to as the modified EfficientNetB0 (m-EfficientNetB0). Our approach targets the MBConv modules in Stage 7 and Stage 8 by reducing their parameter count and computational complexity. Figure 5 illustrates these modifications, with blue text highlighting the reduced modules and dashed lines indicating the removed modules.

Since Stage 7 in the original EfficientNetB0 architecture corresponds to the deep feature extraction phase, having too many stacked layers in this stage can cause feature redundancy. Therefore, we reduced the number of layers in Stage 7 from four to one to mitigate this issue. Regarding Stage 8, which is designed to extract more complex features, if the background in tea leaf disease images is simple, the extracted features may become overly similar, raising the risk of model overfitting. To address this, we removed the MBConv module from Stage 8 entirely. Additionally, we retained the original EfficientNetB0’s end structure, which combines 1 × 1 convolutions, global average pooling, and fully connected layers. This approach preserves key feature representations while significantly reducing computational load and model complexity, resulting in an accurate and efficient deep learning model well-suited for resource-constrained deployment.

2.5. ECA Module

To reduce the number of model parameters and improve inference efficiency, this paper replaces the SE module in the MBConv architecture with the ECA module, enhancing the model’s capacity to identify tea leaf diseases. Figure 6 illustrates the ECA module. Below, we provide an explanation of how the ECA module works.

The ECA module begins by applying global average pooling (GAP) to each channel of the feature map, compressing spatial information into a single representative scalar, as illustrated in Equation (1):

$$z_c={\displaystyle \frac{1}{H\times W}}\sum _{i=1}^H\sum _{j=1}^W{\chi }_c(i,j),$$

(1)

where H and W represent the height and width of the image, respectively. ${\chi }_c$ is the input feature map, and $z_c$ is the average of c channels. This operation generates a channel descriptor vector that highlights the relative importance of each channel while preserving the original number of channels. By focusing on channel-wise significance without dimensionality reduction, the model can more effectively recalibrate feature responses in the subsequent attention mechanism.

Next, the channel descriptor vector $z_c$ is fed into the ECA module, where a one-dimensional convolution is applied to capture local cross-channel dependencies. This convolution enables the model to effectively learn interactions between neighboring channels, enhancing the representation of features without dimensionality reduction. The related concept is formalized in Equation (2):

$${\mathbf{z}}_c^{\prime }=\mathrm{Conv}1\mathrm{D}\left(z_c\right),$$

(2)

where ${\mathbf{z}}_c^{\prime }$ denotes the feature vector. Next, the channel attention weight $s_c$ is generated by applying the sigmoid activation function to the output of the one-dimensional convolution. This sigmoid function produces normalized weights between 0 and 1, reflecting the relative importance of each channel in the feature map. The concept is formalized in Equation (3):

$$s_c=\sigma \left({\mathbf{z}}_c^{\prime }\right),$$

(3)

where $\sigma$ is a sigmoid function. Finally, the channel attention weights $s_c$, generated by the sigmoid function $\sigma$, are multiplied element-wise with the original input feature map ${\chi }_c$ on a channel-by-channel basis. This weighting recalibrates the feature responses to emphasize important channels while suppressing less relevant ones, resulting in the enhanced feature map ${\tilde{\chi }}_c$, as formulated in Equation (4):

$${\tilde{\chi }}_c=s_c⨀{\chi }_c,$$

(4)

where ⨀ denotes element-wise multiplication across channels.

Compared to the SE module, the ECA module enables direct cross-channel interaction through a lightweight one-dimensional convolution without applying dimensionality reduction. This design preserves the complete channel-wise information, avoiding the feature compression and increased computational burden associated with the SE module’s dimensionality reduction and fully connected layers. By maintaining richer channel representations with fewer parameters, the ECA module not only enhances the model’s ability to discriminate subtle features of tea leaf diseases but also significantly reduces computational complexity and resource demands. These advantages align closely with the practical needs of deploying accurate disease recognition models on edge devices with limited memory and processing power, making the ECA module a crucial component in achieving a lightweight, efficient, and effective deep learning solution in agricultural applications.

2.6. ReduceLROnPlateau Learning Rate Adjustment Strategy

Learning rate is a critical training parameter in deep learning models, directly influencing the speed and quality of model convergence. Setting an appropriate learning rate is essential, as it determines how quickly the model updates its parameters and ultimately affects the overall training outcomes. In the context of tea leaf disease image classification, a learning rate that is too high may cause the model to overlook subtle but crucial features such as fine lesions, texture variations, and color differences, leading to suboptimal classification performance.

To address this, we employ the ReduceLROnPlateau learning rate adjustment strategy, which dynamically adjusts the learning rate based on validation loss behavior. Specifically, the strategy monitors the validation loss and reduces the learning rate by a predefined reduction factor (RF) when the loss plateaus—i.e., when no improvement is observed within a specified patience period. This adaptive adjustment enables the model to learn rapidly during the initial training stages while stabilizing the learning process in later stages. Consequently, the model can more effectively capture detailed disease features, enhancing its classification accuracy.

3. Experimental Results

3.1. Experimental Environment

Our experimental setup consists of a 6-core Intel Core i7-8700 CPU, an NVIDIA GeForce RTX 2070 GPU, and 16 GB of RAM. The system runs Windows 11 Professional as the operating system. For software, we use Anaconda version 23.7.4 with Python 3.10.14 as the development environment. The deep learning framework is PyTorch 2.4.1, utilizing CUDA 12.4 and cuDNN 9.1.0 GPU drivers.

To ensure fair comparison across lightweight models and maintain experimental consistency, this study did not include systematic hyperparameter optimization. Rather, we employed widely adopted hyperparameter settings from previous lightweight CNN studies, which have demonstrated effectiveness in similar classification tasks.

Table 1. The various hyperparameter settings used in the proposed approach.

Hyperparameter	Value/Method
Epoch	100
Batch size	16
Learning rate	0.001
Optimizer	Adam
Loss function	Cross-Entropy
Patience in ReduceLROnPlateau	15
Reduction factor in ReduceLROnPlateau	0.5

displays the hyperparameter settings applied in our proposed approach. These parameter settings remained consistent throughout the training phase.

3.2. Image Database of Tea Leaf Diseases

We selected two publicly available image datasets of tea leaf diseases with relatively large numbers of images as experimental datasets to evaluate the performance of the proposed method. These datasets encompass a broad spectrum of tea leaf disease types and feature images captured against simple white backgrounds. The uniform, uncluttered backgrounds minimize extraneous visual interference, allowing the recognition model to concentrate on the distinctive characteristics of the leaf diseases for more accurate classification.

The first dataset is sourced from the Kaggle platform and is named the Tea_Leaf_ Disease [26]. It contains 5867 images of tea leaf diseases, each with a resolution of 256 × 256 pixels. The dataset is split into 80% training-phase data and 20% testing-phase data, corresponding to 4693 and 1174 images, respectively. During the training phase, 10-fold cross-validation is applied to further divide the data into a training set and a validation set, resulting in approximately 4224 training images and 469 validation images.

The second dataset, also obtained from the Kaggle platform, is called the teaLeafBD [27]. It comprises 5276 images of tea leaf diseases, each with a resolution of 1200 × 1594 pixels. To ensure consistent image resolution and eliminate potential bias caused by size differences between the two datasets, all images in this dataset were standardized to the resolution of the Tea_Leaf_Disease dataset prior to training. The data partitioning strategy is the same as that used for the first dataset, dividing the data into 80% training-phase data and 20% testing-phase data, corresponding to 4218 and 1058 images, respectively. During the training phase, 10-fold cross-validation is also applied to further split the training-phase data into a training set and a validation set, resulting in approximately 3797 training images and 421 validation images.

Table 2. Category distribution of tea leaf disease image datasets. The Tea_Leaf_Disease dataset includes six categories, while the teaLeafBD dataset contains seven categories.

Disease Symptom	Tea_Leaf_Disease	teaLeafBD
Algal Spot	1000	418
Brown Blight	867	506
Gray Blight	1000	1013
Healthy	1000	935
Helopeltis	1000	607
Red Spot	1000	NA
Red Spider	NA	515
Green Mirid Bug	NA	1282
Total	5867	5276

Note: “NA” indicates that the corresponding dataset does not include this disease category.

summarizes the disease categories included in the two datasets. To offer a more intuitive understanding of these categories, Figure 7 and Figure 8 showcase sample images from the Tea_Leaf_Disease and teaLeafBD datasets, respectively.

The Tea_Leaf_Disease dataset includes five common tea leaf disease categories and a healthy leaf category: Algal Spot, Brown Blight, Gray Blight, Helopeltis (tea mosquito bug), and Red Spot. Sample images illustrating these conditions are presented in Figure 7. These images visually highlight distinct symptoms such as lesions, spots, and discoloration characteristics of each disease, providing essential visual references to support the classification tasks described.

The teaLeafBD dataset includes six common tea leaf disease categories and a healthy leaf category: Tea Algal Leaf Spot, Brown Blight, Gray Blight, Helopeltis (tea mosquito bug), Red Spider, Green Mirid Bug, and healthy leaf. Sample images illustrating these conditions are shown in Figure 8. These images highlight distinctive symptoms such as spots, lesions, discoloration, and pest damage characteristics of each disease, providing valuable visual references that support the classification and identification tasks.

While the category distributions in Table 1 result from the data collection and annotation processes—and may not precisely reflect the actual ratios of healthy to diseased tea leaves in real tea gardens—this study primarily uses these public datasets as benchmarks to validate and compare the classification performance of the proposed model. Overall, these two tea leaf disease datasets encompass diseases caused by diverse pathogens such as fungi, viruses, and insects, highlighting the rich variations in morphology, color, and texture of tea leaf diseases. This complexity enables a more robust evaluation and comparison of different recognition methods.

3.3. Experimental Results and Analysis

In this section, we present five experiments designed to evaluate the effectiveness of the proposed methodology. All experiments utilize the Tea_Leaf_Disease and teaLeafBD datasets for tea leaf disease identification.

• Experiment 1 investigates the impact of data augmentation strategies on model performance.
• Experiment 2 compares the original EfficientNetB0 architecture with an m-EfficientNetB0 module to select the optimal architecture.
• Experiment 3 builds upon the best architecture identified in Experiment 2 to explore the influence of various attention modules on model recognition performance.
• Experiment 4 examines the effect of different learning rate adjustment strategies on model accuracy.
• Experiment 5 compares the proposed model against other lightweight models to verify its effectiveness.

This study utilizes several evaluation metrics—accuracy, precision, recall, and F1-score—to thoroughly assess the classification model’s effectiveness. Accuracy serves as the main metric to gauge the overall correctness of the model’s predictions per class. Precision evaluates the model’s ability to minimize false positives by avoiding the mislabeling of other classes as the target class, while recall measures the model’s success in detecting actual instances of the class without missing them. The F1-score [28] is used to balance precision and recall by calculating their harmonic mean, offering an informative and balanced metric—especially important when both false positives and false negatives have significant consequences in disease classification. Thus, the F1-score serves as a key evaluation metric that combines precision and recall into a single measure. Considering the potential imbalance of sample sizes across classes in this multi-class setting, the macro averaging approach [28] is adopted to ensure equal weight for each category. This approach helps prevent the model from performing well on one metric while neglecting the other, thus offering a more balanced and comprehensive performance evaluation. The formal definitions of accuracy, precision, recall, and F1-score for the multi-class classification problem are given in Equations (5)–(8).

These metrics are calculated based on the following definitions:

• True positive (TP) refers to actual positive samples correctly predicted as positive.
• False positive (FP) denotes actual negative samples incorrectly predicted as positive.
• False negative (FN) indicates actual positive samples incorrectly predicted as negative.
• True negative (TN) represents actual negative samples correctly predicted as negative.

$$Accurac{y}_{marco}={\displaystyle \frac{1}{N}}\sum _{i=1}^N{\displaystyle \frac{T{P}_i+T{N}_i}{T{P}_i+T{N}_i+F{P}_i+F{N}_i}}$$

(5)

$$Precisio{n}_{marco}={\displaystyle \frac{1}{N}}\sum _{i=1}^N{\displaystyle \frac{T{P}_i}{T{P}_i+F{P}_i}}$$

(6)

$$Recal{l}_{marco}={\displaystyle \frac{1}{N}}\sum _{i=1}^N{\displaystyle \frac{T{P}_i}{T{P}_i+F{N}_i}}$$

(7)

$$F1-scor{e}_{marco}={\displaystyle \frac{1}{N}}\sum _{i=1}^N{\displaystyle \frac{2\times Precisio{n}_i\times Recal{l}_i}{Precisio{n}_i+Recal{l}_i}}$$

(8)

where i denotes the i-th class, and N represents the total number of classes.

3.3.1. Impact of Image Enhancement on the Identification Performance

To improve the performance of the original EfficientNetB0 model during training, we employed various data augmentation techniques, including vertical (up–down) flip, horizontal (left–right) flip, random rotation, random crop, random affine transformation, and random brightness adjustment. The settings for these image enhancement techniques are shown in

Table 3. Image enhancement techniques used for the training phase.

Image Enhancement	Detailed Description
Random vertical and horizontal flip	Each applied with a 50% probability
Random rotation	$\pm {20}^{\circ }$ range
Random crop	80–100% of original area
Random affine transformation	Random translation 0–10% of image size
Random brightness adjustment	$\pm 20\%$ range

.

These methods effectively simulate the variability encountered in actual farm environments, such as diverse distributions of tea leaf spots and fluctuating lighting conditions. By increasing the diversity of training samples through these transformations, the model’s generalization ability is enhanced, helping it better adapt to different real-world scenarios. The performance results of applying image augmentation techniques on the two tea leaf image datasets are presented in

Table 4. Average 10-fold cross-validation accuracy and standard deviation with/without image enhancement on the Tea_Leaf_Disease dataset during the training phase.

Image Enhancement	Average Accuracy (%)	Std. Dev.
No	98.51	0.32
Yes	99.57	0.39

,

Table 5. EfficientNetB0 performance on the Tea_Leaf_Disease testing set with/without image enhancement.

Image Enhancement	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)
No	98.21	98.20	98.15	98.17
Yes	99.23	99.21	99.24	99.22

,

Table 6. Average 10-fold cross-validation accuracy and standard deviation with/without image enhancement on the teaLeafBD dataset during the training phase.

Image Enhancement	Average Accuracy (%)	Std. Dev.
No	81.36	1.89
Yes	91.94	1.33

and

Table 7. EfficientNetB0 performance on the teaLeafBD testing set with/without image enhancement.

Image Enhancement	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)
No	77.50	75.13	74.55	74.84
Yes	88.56	87.68	85.93	86.80

.

From the experimental results in Table 4, it is evident that applying data augmentation strategies significantly enhances the average classification accuracy of the EfficientNetB0 model under 10-fold cross-validation while maintaining strong stability.

The experimental results, as shown in Table 5, demonstrate that applying a data augmentation strategy significantly improves the performance of the EfficientNetB0 model by more than 1 percentage point across all evaluated metrics, with a particularly notable increase in recall. This enhancement indicates that data augmentation effectively boosts the model’s classification accuracy and reduces false negatives, thereby making it more robust and suitable for the variable and complex environmental conditions encountered in farmland.

The experimental results in Table 6 clearly show that employing data augmentation strategies significantly improves the average accuracy of the EfficientNetB0 model while reducing its standard deviation. This indicates that data augmentation not only enhances the model’s recognition performance but also provides better classification stability.

Experimental results in Table 7 indicate that applying image augmentation techniques significantly improves all evaluation metrics, with the greatest increase observed in precision. This demonstrates that image augmentation notably enhances the model’s ability to correctly classify specific diseases by effectively reducing the misclassification of other categories as the target class. Overall, data augmentation can substantially boost the model’s classification performance on disease images for certain datasets.

3.3.2. Performance Comparison of Different EfficientNetB0 Architectures

To verify the effectiveness of the proposed architectural improvement strategies, we conducted a comparative analysis of four architecture variants based on the original EfficientNetB0 model: (1) the baseline EfficientNetB0, (2) the model with a purely modified Stage7 module (denoted as EfficientNetB0 ⊕ Red. S7), (3) the model with a purely removed Stage8 module (denoted as EfficientNetB0 − S8), and (4) the fully improved EfficientNetB0 model (denoted as m-EfficientNetB0) that integrates both Stage 7 and Stage 8 enhancements.

These variants were comprehensively evaluated by measuring the average classification accuracy and standard deviation using 10-fold cross-validation. Relevant performance metrics including accuracy, training time, inference time, and the number of model parameters were also selected for assessment. The analysis of model complexity and runtime efficiency further validates the practical advantages gained from the improved design. The detailed experimental results are presented in

Table 8. Comparison of architectures based on average 10-fold cross-validation accuracy and standard deviation on the Tea_Leaf_Disease dataset during the training phase.

Architecture	Average Accuracy (%)	Std. Dev.	Training Time (s)
EfficientNetB0	99.57	0.39	49,763
EfficientNetB0 ⊕ Red. S7	99.64	0.15	46,426
EfficientNetB0 − S8	99.66	0.30	48,048
m-EfficientNetB0	99.81	0.19	45,709

,

Table 9. EfficientNetB0 variant comparison on the Tea_Leaf_Disease testing set.

Architecture	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	Inference Time (s)
EfficientNetB0	99.23	99.21	99.24	99.22	0.0170
EfficientNetB0 ⊕ Red. S7	99.23	99.20	99.23	99.21	0.0119
EfficientNetB0 − S8	99.32	99.29	99.33	99.31	0.0136
m-EfficientNetB0	99.40	99.37	99.40	99.38	0.0113

,

Table 10. Comparison of architectures based on average 10-fold cross-validation accuracy and standard deviation on the teaLeafBD dataset during the training phase.

Architecture	Average Accuracy (%)	Std. Dev.	Training Time (s)
EfficientNetB0	91.94	1.33	48,219
EfficientNetB0 ⊕ Red. S7	92.34	1.09	45,302
EfficientNetB0 − S8	91.92	0.90	47,057
m-EfficientNetB0	91.87	1.40	44,476

,

Table 11. EfficientNetB0 variant comparison on the teaLeafBD testing set.

Architecture	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	Inference Time (s)
EfficientNetB0	88.56	87.68	85.93	86.80	0.0171
EfficientNetB0 ⊕ Red. S7	88.37	86.67	87.55	87.11	0.0122
EfficientNetB0 − S8	88.37	86.67	86.37	86.52	0.0136
m-EfficientNetB0	89.32	87.33	86.61	87.07	0.0114

and

Table 12. Model complexity of different EfficientNetB0 variants.

Architecture	GFLOPs	Size (MB)	Num. of Parameters
EfficientNetB0	0.413874	15.60	4,016,515
EfficientNetB0 ⊕ Red. S7	0.342640	8.79	2,252,659
EfficientNetB0 − S8	0.370807	11.82	3,032,531
m-EfficientNetB0	0.299574	5.01	1,268,675

.

Table 8 shows that the proposed m-EfficientNetB0 architecture achieves the highest average accuracy with a lower standard deviation on the Tea_Leaf_Disease dataset. This indicates its high performance and reliable consistency. Additionally, this architecture has the shortest training time, demonstrating excellent computational efficiency. These results show that the multi-stage optimization strategy adopted by m-EfficientNetB0 effectively balances high accuracy and computational efficiency, making it the best choice for tea leaf disease classification tasks.

According to Table 9, optimizing only Stage 7 lowers inference time, indicating that enhancements to Stage 7 contribute to greater computational efficiency. Modifications to Stage 8 further reduce inference time and also improve all evaluation metrics, which suggests strengthened feature extraction capabilities. The proposed m-EfficientNetB0 model, which integrates improvements to both Stage 7 and Stage 8, delivers a balanced architecture that not only promotes computational efficiency but also achieves superior performance in accuracy, precision, recall, and F1-score compared to single-stage modifications.

From Table 10, we observed that while m-EfficientNetB0 has a slightly lower average accuracy on the teaLeafBD dataset compared to other architectures, its overall performance is still competitive. The model remains robust and stable, effectively recognizing subtle differences in high-resolution tea leaf images, which is valuable for precision disease diagnosis. This consistency makes m-EfficientNetB0 a practical and reliable choice for challenging image recognition tasks in tea leaf diseases identification.

The experimental results in Table 11 demonstrate that the m-EfficientNetB0 architecture outperforms both the original model and the single-module variants in terms of accuracy, indicating that the two-stage modular improvement strategy effectively complements each stage to enhance the model’s capacity for disease feature classification. Additionally, the significant reduction in inference time confirms that the proposed m-EfficientNetB0 achieves a favorable balance between enhanced accuracy and computational efficiency.

The experimental results in Table 12 indicate that the single-stage improvements in both Stage 7 and Stage 8 contribute to a moderate reduction in computational cost and model size. However, our m-EfficientNetB0 architecture achieves a more substantial reduction in computation, model size, and number of parameters. Compared to the original EfficientNetB0, the GFLOPs decrease from approximately 0.4139 to 0.2996, the model size is reduced from 15.6 MB to 5.01 MB, and the number of parameters is reduced from around 4 million to 1.26 million. These results demonstrate that the m-EfficientNetB0 effectively lowers resource consumption and is better suited for deployment on edge devices.

3.3.3. Performance Comparison of m-EfficientNetB0 with Different Attention Modules

To assess the impact of different attention modules on the proposed method, we conducted an experiment in which the SE attention module originally used in EfficientNetB0 was replaced with alternative attention mechanisms. These modules were integrated into the modified EfficientNetB0 architecture employed in our study. The attention modules selected for comparison include the Convolutional Block Attention Module (CBAM), Coordinate Attention (CA), Multidimensional Collaborative Attention (MCA) [29], Shuffle Attention (SA) [30], and the ECA module. The corresponding experimental results are presented in

Table 13. Average 10-fold cross-validation accuracy and standard deviation of our method with various attention modules on the Tea_Leaf_Disease dataset during the training phase.

Method	Average Accuracy (%)	Std. Dev.	Training Time (s)
m-EfficientNetB0	99.81	0.19	45,709
m-EfficientNetB0 + CBAM	98.96	0.63	51,219
m-EfficientNetB0 + CA	99.60	0.32	49,533
m-EfficientNetB0 + MCA	99.70	0.18	59,521
m-EfficientNetB0 + SA	99.57	0.30	49,330
m-EfficientNetB0 + ECA	99.81	0.21	47,353

,

Table 14. Comparison of m-EfficientNetB0 with different attention modules on the Tea_Leaf_Disease testing set.

Method	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	Inference Time (s)
m-EfficientNetB0	99.40	99.37	99.40	99.38	0.0113
m-EfficientNetB0 + CBAM	98.38	98.33	98.38	98.35	0.0187
m-EfficientNetB0 + CA	99.06	99.04	99.08	99.06	0.0160
m-EfficientNetB0 + MCA	98.89	98.84	98.92	98.88	0.0317
m-EfficientNetB0 + SA	98.98	98.92	98.99	98.95	0.0163
m-EfficientNetB0 + ECA	99.49	99.49	99.48	99.48	0.0098

,

Table 15. Average 10-fold cross-validation accuracy and standard deviation of our method with various attention modules on the teaLeafBD dataset during the training phase.

Method	Average Accuracy (%)	Std. Dev.	Training Time (s)
m-EfficientNetB0	91.87	1.40	44,476
m-EfficientNetB0 + CBAM	89.97	1.61	48,414
m-EfficientNetB0 + CA	92.25	1.09	48,864
m-EfficientNetB0 + MCA	92.20	1.02	56,191
m-EfficientNetB0 + SA	92.27	1.82	47,141
m-EfficientNetB0 + ECA	92.27	1.54	45,343

,

Table 16. Comparison of m-EfficientNetB0 with different attention modules on the teaLeafBD testing set.

Method	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	Inference Time (s)
m-EfficientNetB0	89.32	87.33	86.61	87.07	0.0114
m-EfficientNetB0 + CBAM	88.85	88.02	86.05	87.02	0.0182
m-EfficientNetB0 + CA	88.85	87.37	86.58	86.97	0.0142
m-EfficientNetB0 + MCA	88.75	86.77	88.36	87.56	0.0268
m-EfficientNetB0 + SA	89.51	87.88	88.51	88.19	0.0138
m-EfficientNetB0 + ECA	90.73	89.97	88.51	89.23	0.0095

and

Table 17. Model complexity comparison of m-EfficientNetB0 with different attention modules.

Method	GFLOPs	Size (MB)	Num. of Parameters
m-EfficientNetB0	0.299574	5.01	1,268,675
m-EfficientNetB0 + CBAM	0.300331	5.34	1,356,127
m-EfficientNetB0 + CA	0.304910	5.89	1,494,759
m-EfficientNetB0 + MCA	0.303574	4.30	1,079,427
m-EfficientNetB0 + SA	0.298346	4.30	1,079,303
m-EfficientNetB0 + ECA	0.299402	4.27	1,079,339

.

Table 13 demonstrates that m-EfficientNetB0 offers the optimal trade-off, providing leading accuracy and the shortest training duration with minimal variability. Compared to the baseline, most attention modules increase computational demands without yielding notable accuracy gains. CA, MCA, and SA produce similar results to the baseline but are less efficient, while CBAM delivers both lower accuracy and higher variance. Among all evaluated modules, only ECA matches the baseline accuracy and maintains efficient training. Given ECA’s superior inference efficiency, it is selected as the replacement for the SE attention module.

The experimental results in Table 14 show that our proposed method delivers the best performance, with significant improvements in both accuracy and efficiency. While attention modules such as CA, MCA, and SA achieve reasonable results, they suffer from slower inference speeds. In contrast, CBAM performs the worst overall. Consequently, m-EfficientNetB0 paired with the ECA module emerges as the most effective configuration for accurately identifying tea leaf diseases.

As shown in Table 15, the baseline m-EfficientNetB0 achieves strong accuracy with the shortest training duration. Attention modules provide modest accuracy enhancements, with CA and SA/ECA leading in performance. MCA offers the most stable outcomes, reflected in the lowest standard deviation, but incurs the longest training time. CBAM performs the weakest, showing the lowest accuracy and greatest variability. Overall, attention modules improve accuracy slightly but increase training times. Notably, ECA integration boosts average accuracy significantly while reducing deviation, highlighting its effectiveness in enhancing both model performance and stability during classification.

The experimental results presented in Table 16 indicate that the m-EfficientNetB0 model with the ECA module achieves the best results. Compared to other attention modules, the ECA module introduces a slight increase in training time but significantly enhances performance, underscoring the model’s strong adaptability on the teaLeafBD dataset. This suggests that the ECA module serves as a highly effective enhancement for improving model accuracy and efficiency.

The experimental results in Table 17 indicate that the integration of the ECA module with m-EfficientNetB0 indeed achieves the most lightweight design among the various attention modules tested. Compared to more complex modules such as SE and CA, the ECA module substantially reduces both model size and the number of parameters, striking a balance between computational cost and architectural simplicity. This demonstrates the inherent advantage of the ECA module in terms of lightweight architecture.

3.3.4. Performance Comparison of Applying Learning Rate Adjustment

In this experiment, we employ the modified EfficientNetB0 combined with the ECA module to examine the impact of the learning rate adjustment strategy on model performance. Specifically, we utilize the ReduceLROnPlateau strategy with monitoring set to verify loss minimization. To prevent premature reductions in the learning rate that could hinder potential improvements, we set the patience parameter to 15, acknowledging that steady performance gains typically occur within the initial rounds of training. Additionally, the reduction factor is set to 0.5, halving the learning rate at each adjustment step to facilitate finer weight updates during convergence. This approach ultimately aims to enhance classification performance. The corresponding experimental results are presented in

Table 18. Average 10-fold cross-validation accuracy and standard deviation with/without ReduceLROnPlateau on the Tea_Leaf_Disease dataset during the training phase.

ReduceLROnPlateau	Average Accuracy (%)	Std. Dev.
No	99.47	0.27
Yes	99.81	0.19

,

Table 19. Effect of ReduceLROnPlateau on the Tea_Leaf_Disease testing set.

ReduceLROnPlateau	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)
No	98.81	98.79	98.82	98.80
Yes	99.49	99.49	99.48	99.48

,

Table 20. Average 10-fold cross-validation accuracy and standard deviation with/without ReduceLROnPlateau on the teaLeafBD dataset during the training phase.

ReduceLROnPlateau	Average Accuracy (%)	Std. Dev.
No	91.75	1.41
Yes	92.27	1.54

and

Table 21. Effect of ReduceLROnPlateau on the teaLeafBD testing set.

ReduceLROnPlateau	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)
No	87.24	86.12	85.39	85.75
Yes	90.73	89.97	88.51	89.23

.

As shown in Table 18, it is clear that using the ReduceLROnPlateau learning rate adjustment strategy significantly improves average accuracy while reducing standard deviation. This indicates that the strategy not only enhances overall classification performance but also enables more stable model convergence during training.

As shown in Table 19, incorporating the ReduceLROnPlateau learning rate adjustment strategy effectively enhances the model’s performance in disease recognition. All evaluated metrics exhibit an upward trend, demonstrating that this adjustment strategy positively contributes to the overall effectiveness of the model in accurately identifying diseases.

Based on the experimental results in Table 20, although the standard deviation slightly increases when using the ReduceLROnPlateau learning rate adjustment strategy, the model’s average accuracy still shows improvement. This indicates that the strategy enhances the overall classification performance.

Table 21 shows that using the ReduceLROnPlateau learning rate adjustment strategy significantly boosts accuracy on the teaLeafBD dataset. Furthermore, precision, recall, and F1-Score metrics all increase noticeably. These results highlight the crucial role of learning rate adjustment in enhancing the model’s recognition accuracy.

3.3.5. Performance Comparison with Other Lightweight Models

To evaluate the performance of the proposed method in the context of lightweight models, we conducted a comparative analysis against several representative lightweight deep learning architectures, each with computational requirements below 1 GFLOP per inference. The baseline models selected for comparison include SqueezeNet 1.1 [31], MNASNet1_3 [32], ShuffleNetV2_x2_0 [33], MobileNetV4_conv_small [34], and RepGhostNet1.0× [35]. For fairness, all lightweight networks compared were configured with exactly the same hyperparameters as our method. Detailed experimental results are provided in

Table 22. Average 10-fold cross-validation accuracy and standard deviation of different lightweight models on the Tea_Leaf_Disease dataset during the training phase.

Method	Average Accuracy (%)	Std. Dev.	Training Time (s)
SqueezeNet 1.1	97.59	1.07	37,356
MNASNet1_3	99.11	0.31	48,445
ShuffleNetV2_x2_0	99.42	0.35	43,440
MobileNetV4_conv_small	98.79	0.38	43,722
RepGhostNet 1.0×	99.42	0.34	48,627
Ours	99.81	0.19	47,353

,

Table 23. Performance comparison of different lightweight models on the Tea_Leaf_Disease testing set.

Method	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	Inference Time (s)
SqueezeNet 1.1	95.82	95.80	95.88	95.84	0.0044
MNASNet1_3	99.06	99.08	99.07	99.07	0.0099
ShuffleNetV2_x2_0	98.98	98.94	98.99	98.96	0.0113
MobileNetV4_conv_small	97.53	97.53	97.50	97.51	0.0100
RepGhostNet 1.0×	99.15	99.14	99.15	99.14	0.0186
Ours	99.49	99.49	99.48	99.48	0.0098

,

Table 24. Average 10-fold cross-validation accuracy and standard deviation of different lightweight models on the teaLeafBD dataset during the training phase.

Method	Average Accuracy (%)	Std. Dev.	Training Time (s)
SqueezeNet 1.1	84.19	1.70	38,939
MNASNet1_3	88.24	1.81	46,543
ShuffleNetV2_x2_0	90.30	1.59	43,088
MobileNetV4_conv_small	87.48	1.26	42,385
RepGhostNet 1.0×	90.30	1.34	49,058
Ours	92.27	1.54	45,343

,

Table 25. Performance comparison of different lightweight models on the teaLeafBD testing set.

Method	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	Inference Time (s)
SqueezeNet 1.1	79.49	80.82	73.74	77.12	0.0049
MNASNet1_3	83.84	83.45	80.64	82.02	0.0114
ShuffleNetV2_x2_0	88.28	86.73	86.02	86.37	0.0147
MobileNetV4_conv_small	84.97	83.60	83.06	83.33	0.0099
RepGhostNet 1.0×	88.00	87.40	85.70	86.54	0.0205
Ours	90.73	89.97	88.51	89.23	0.0095

and

Table 26. Model complexity comparison of different lightweight models.

Method	GFLOPs	Size (MB)	Num. of Parameters
SqueezeNet 1.1	0.263231	2.78	726,087
MNASNet1_3	0.554501	19.40	5,010,223
ShuffleNetV2_x2_0	0.596196	20.68	5,359,339
MobileNetV4_conv_small	0.184783	9.74	2,476,903
RepGhostNet 1.0×	0.164001	11.05	2,801,571
Ours	0.299402	4.27	1,079,339

.

The experimental findings in Table 22 indicate that the proposed approach achieves the highest average accuracy along with the lowest standard deviation, demonstrating its robust consistency. In comparison to other lightweight models, it provides comparable training durations. These findings illustrate the proposed method’s dependability and efficiency, establishing it as a favored option for precise and consistent disease classification in tea leaf datasets.

According to the experimental results in Table 23, it is evident that our method does not achieve the shortest inference time, as SqueezeNet 1.1 records a lower value in this metric. Nonetheless, our approach excels by delivering the highest accuracy, precision, recall, and F1-Score on the Tea_Leaf_Disease dataset, outperforming lightweight models such as SqueezeNet 1.1 and MobileNetV4_conv_small. These results emphasize the superior accuracy of our method, establishing it as a highly effective solution for tea leaf disease detection, notwithstanding the inference time consideration.

The examination of Table 24 indicates that, even though our approach needs a longer training duration than SqueezeNet 1.1, MobileNetV4_conv_small, and ShuffleNetV2_x2_0, it attains the best average accuracy on the teaLeafBD dataset. This is backed by a low standard deviation, signifying steady performance. The enhanced precision and generalization ability of our method clearly indicate its success in boosting model performance.

Table 25 shows that our method achieves the highest accuracy, precision, recall, and F1-Score on the teaLeafBD dataset, outperforming all lightweight network models used for comparison. Although it does not have the shortest inference time, its strong performance metrics highlight its effectiveness for tea leaf disease classification. These results emphasize the robustness of our approach in delivering superior classification accuracy.

Table 26 presents a comprehensive comparison of model complexity across various lightweight architectures. Our approach achieves the lowest GFLOPs, indicating lower computational complexity relative to alternatives such as MNASNet1_3 and ShuffleNetV2_x2_0. This effectiveness is also demonstrated by the compact model size and the least parameters, significantly outperforming MNASNet1_3 and ShuffleNetV2_x2_0. Our approach clearly shows efficient resource utilization. Its straightforward yet effective design makes it ideal for deployment in resource-constrained environments such as embedded systems or mobile devices, where both efficacy and efficiency are important. These results validate that our method consistently achieves a reliable balance between computational efficiency and robust model performance.

3.4. Discussion

To gain deeper insights into the model’s classification behavior across different disease categories, the classification results of the two tea leaf disease testing sets are presented as confusion matrices. In these matrices, the horizontal axis represents the predicted labels, while the vertical axis indicates the true labels. Diagonal elements correspond to correctly classified samples, and off-diagonal elements highlight misclassifications between categories. Analyzing these confusion matrices enables us to better evaluate our method’s effectiveness on the two tea leaf disease datasets and investigate the potential causes of classification errors.

Figure 9 shows the confusion matrix obtained using the testing set from the Tea_Leaf_Disease dataset. As illustrated in the confusion matrix, our method achieves high accuracy in identifying Algal Spot disease and Gray Blight disease, with no observed misclassifications. However, Brown Blight exhibited the highest rate of misclassification, with two instances incorrectly identified as Gray Blight, as shown in Figure 10. It is hypothesized that the observed misclassification primarily results from the high degree of visual similarity in symptom presentation between the two diseases. Particularly, the manifestation of dark brown spots during the later stages introduces considerable ambiguity in the input data, thereby limiting the model’s ability to effectively distinguish between these classes based on visual features alone.

Figure 11 shows the confusion matrix obtained using the testing set from the teaLeafBD dataset. As shown in the confusion matrix, our method demonstrates robust classification performance. However, there are notable confusions between certain categories. Specifically, 14 images of Brown Blight were misclassified as Tea Algal Leaf Spot, and 11 as Gray Blight, while 14 images of Red Spider were also misclassified as Gray Blight. These misclassifications suggest the presence of highly similar features among these disease categories, which likely contribute to the model’s errors.

The misclassified image samples from the teaLeafBD dataset, presented in Figure 12, highlight the challenges in distinguishing visually similar diseases. Figure 12a shows a correctly classified Tea Algal Leaf Spot, characterized by bright yellow or orange spots. In contrast, Figure 12b depicts Brown Blight, which initially presents as yellow-colored, round or irregular spots closely resembling Tea Algal Leaf Spot, leading to misclassification by the model. Figure 12c illustrates Gray Blight, identified by its distinctive brown wheel-shaped spots. The misclassification of severe Brown Blight cases (Figure 12d) as Gray Blight is likely due to the dark brown coloration developing in later disease stages, creating visual similarities that confuse the model. These findings underscore the intrinsic difficulty in classifying diseases with overlapping symptom presentations and color patterns from leaf images, a common challenge documented in tea leaf disease recognition research.

Figure 13 illustrates a misclassification instance wherein our proposed approach incorrectly identified symptoms of Red Spider as those of Gray Blight. Specifically, Figure 13a displays a correctly classified Gray Blight image, whereas Figure 13b, despite depicting Red Spider symptoms, was erroneously classified as Gray Blight by the model. A detailed visual comparison reveals that both images contain irregular brown lesions, a feature that likely contributed to the model’s confusion during classification.

We hypothesize that the misclassification arises because Red Spider feeding commonly results in dried spots on the leaf, which closely resemble the fungal lesions characteristic of Gray Blight, thus creating overlapping visual features. Moreover, authentic Red Spider symptoms are generally accompanied by subtle puncture marks and webbing on the leaf surface; however, these distinguishing characteristics were either absent or not prominently visible in the images utilized. Consequently, the model predominantly relies on the lesion’s morphological attributes, such as shape and coloration, thereby leading to the incorrect categorization of Red Spider symptoms as Gray Blight.

The differences in model performance observed between the two tea leaf disease datasets, although the tasks seem similarly complex, can be attributed to intrinsic factors related to the datasets and experimental conditions.

• Although the number of disease categories is similar, the specific classes vary between the datasets, with some diseases unique to each. The teaLeafBD dataset includes additional categories such as Green Mirid Bug and Red Spider, which may be less visually distinct or more difficult to differentiate compared to those in the Tea_Leaf_Disease dataset. Furthermore, differences in the sample size per category (class imbalance) can also impact model training efficacy and prediction accuracy.
• Both datasets feature predominantly plain white backgrounds that reduce noise; however, subtle variations in lighting, leaf orientation, and leaf health conditions may still occur. These differences in image acquisition parameters can impact the model’s generalization capability and may explain the observed performance discrepancies.
• The morphological characteristics of tea leaves—such as lesion shape, size, and distribution—together with textural variations including roughness and speckled lesion patterns, constitute essential indicators for disease detection. The two tea disease datasets exhibit significant morphological and textural differences, which in turn affect the model’s performance on each dataset.

Differences in category subtlety, sample distribution, and visual complexity largely explain the performance variance between the two datasets. Although the classification tasks appear comparable superficially, these intrinsic factors influence the model’s effectiveness significantly, emphasizing the importance of dataset characteristics in performance interpretation. This analysis supports careful selection and benchmarking across diverse datasets for robust model validation.

4. Conclusions

Deep learning techniques have demonstrated strong potential in plant disease recognition. However, achieving high accuracy while reducing the computational demands of models remains a significant and pressing challenge. To address this issue, this study proposes a tea leaf disease identification model that balances high recognition accuracy with low computational costs. By improving the Stage 7 and Stage 8 modules of EfficientNetB0, we effectively reduce computational complexity and parameter count. Experimental results on two publicly available tea leaf disease datasets show that our method outperforms other lightweight models across multiple evaluation metrics, such as accuracy and precision, highlighting its potential applicability in field disease monitoring and real-time diagnosis. Nevertheless, since the datasets used in this paper consist of single-leaf images captured against plain backgrounds, there is still a gap compared to real field conditions. Future studies will require validation using more in-field image data to assess practical applicability.

Future research will focus on several directions. First, we will expand the dataset by collecting more real-world field images to retrain and fine-tune the model, further verifying and strengthening its performance in practical scenarios. In addition, we plan to perform multiple rounds of random partitioning and statistical analysis to quantify the influence of different data splits on model performance variability, thereby establishing a more comprehensive and reliable evaluation framework. Second, we plan to broaden data enhancement strategies to better simulate variations encountered during image acquisition, such as shadows caused by sunlight, overlapping leaves, leaf edges touching the image boundary, and motion blur. Incorporating these factors is expected to enhance the generalization ability of our method under complex field conditions. In addition, we aim to integrate multi-scale feature learning strategies to improve the model’s ability to distinguish different disease patterns and subtle feature variations. Finally, to gain a more comprehensive understanding of the characteristic features of tea leaf diseases, we will perform visualization analyses, employing techniques such as Gradient-Weighted Class Activation Mapping (Grad-CAM), to elucidate the model’s reliance on various image features. Furthermore, the model will be deployed on edge computing devices, including smartphones and drones, facilitating real-time diagnosis and thereby contributing to the advancement of automated smart agriculture systems. This deployment aims to enhance operational efficiency and enable timely disease management in agricultural settings.