TOM-SSL: Tomato Disease Recognition Using Pseudo-Labelling-Based Semi-Supervised Learning

Abstract

In the agricultural domain, the availability of labelled data for disease recognition tasks is often limited due to the cost and expertise required for annotation. In this paper, a novel semi-supervised learning framework named TOM-SSL is proposed for automatic tomato leaf disease recognition using pseudo-labelling. TOM-SSL effectively addresses the challenge of limited labelled data by leveraging a small labelled subset and confidently pseudo-labelled samples from a large pool of unlabelled data to improve classification performance. Utilising only 10% of the labelled data, the proposed framework with a MobileNetV3-Small backbone achieves the best accuracy at 72.51% on the tomato subset of the PlantVillage dataset and 70.87% on the Taiwan tomato leaf disease dataset across 10 disease categories in PlantVillage and 6 in the Taiwan dataset. While achieving recognition performance on par with current state-of-the-art supervised methods, notably, the proposed approach offers a tenfold enhancement in label efficiency.

1. Introduction

Feeding the projected global population of 10 billion by 2050 will require a substantial increase in food production, estimated at around 60% above current levels [1]. This demand places immense pressure on the agricultural sector, which is already constrained by numerous challenges, including climate change, land degradation and the prevalence of crop diseases. Tomatoes, a globally important vegetable crop valued for their nutritional and economic significance, are particularly vulnerable to a wide range of leaf diseases caused by fungal, bacterial and viral pathogens [2,3]. These diseases not only diminish crop yield but also degrade fruit quality, which threatens the stability of tomato supply chains. According to recent estimates, plant diseases are responsible for considerable yield reductions in staple crops, highlighting the urgent need for scalable and accurate disease detection systems [4]. Early identification of tomato leaf diseases is thus vital for minimising crop loss and improving disease management practices in precision agriculture.

Despite the increasing focus on automated plant disease recognition, a major challenge in the agricultural domain remains the limited availability of labelled data [5]. Acquiring labelled images of diseased crops typically requires expert annotation, which is both time-consuming and costly. In contrast, large volumes of unlabelled plant images can be easily collected using field sensors, drones or mobile devices. This imbalance presents a unique opportunity for semi-supervised learning (SSL), a paradigm that aims to train models using a small amount of labelled data alongside a large pool of unlabelled data [6]. Effectively leveraging limited supervision and abundant unlabelled samples has been a long-standing challenge in machine learning. SSL is increasingly recognised as a promising approach for reducing dependence on manual labelling while maintaining performance under data scarcity. Recently, semi-supervised few-shot learning has been applied in plant disease recognition and has demonstrated promising performance [7].

A variety of SSL techniques have been proposed over the past decades, including consistency regularisation, entropy minimisation, graph-based methods, and generative models [8]. Among these, pseudo-labelling has emerged as one of the most practical and effective approaches, particularly for image classification tasks [9]. It operates by assigning artificial labels to unlabelled data using a model trained on a small labelled subset, enabling the model to iteratively improve its performance as more pseudo-labelled samples are incorporated. Compared to other SSL methods, pseudo-labelling is straightforward to implement, scalable to large datasets, and compatible with existing supervised architectures, making it highly suitable for real-world agricultural applications. Despite its advantages, to the best of the current knowledge, no prior work has applied pseudo-labelling-based semi-supervised learning to the task of tomato leaf disease recognition. This highlights a critical research gap and highlights the need for a data-efficient framework tailored to the unique challenges of tomato leaf disease detection.

A data-efficient pseudo-labelling strategy is introduced through the TOM-SSL framework, which leverages unlabelled samples via an SSL paradigm and employs MobileNetV3-Small as a lightweight backbone to enable early and accurate recognition of tomato leaf diseases under limited labelled data conditions. The key contributions of this paper are given as follows:

• TOM-SSL Framework: A novel pseudo-labelling-based semi-supervised learning framework, as illustrated in Figure 1, is proposed to improve early-stage recognition of diverse tomato leaf diseases.
• Data-Efficient Learning: The proposed TOM-SSL framework addresses the prevalent challenge of limited labelled data in tomato disease recognition by systematically leveraging the abundance of unlabelled data through a semi-supervised learning approach. To further reduce the initial labelled data, MobileNetV3-Small is used as the backbone.
• Extensive Benchmarking: A comprehensive evaluation of the proposed TOM-SSL framework is conducted on two challenging and publicly available tomato leaf datasets (https://data.mendeley.com/datasets/ngdgg79rzb/1, accessed on 1 July 2025). Experimental results demonstrate the superior performance of TOM-SSL in accurately recognising tomato leaf diseases compared to existing state-of-the-art methods under limited labelled data.

Below is the outline for the rest of the paper. Section 2 reviews relevant literature for tomato leaf disease recognition. The proposed TOM-SSL framework used for recognising tomato leaf disease is presented in Section 3. In Section 4, the experiments and the results obtained are presented. Finally, in Section 5, the paper is wrapped up with a conclusion and potential future works.

2. Related Work

Tomato leaf disease classification has been the subject of extensive research over the past decades, reflecting its significance in agricultural health monitoring. This section adopts a two-pronged focus: the first examines general approaches to tomato leaf disease recognition (Section 2.1), while the second reviews methodologies specifically designed to address the challenge of limited labelled data in agriculture (Section 2.2). Readers are encouraged to refer to recent surveys [10,11,12] for a broader overview of the literature on plant disease recognition.

2.1. Tomato Disease Recognition

Early tomato disease recognition studies primarily relied on handcrafted feature extraction and classical machine learning algorithms, such as support vector machines (SVMs) [13], K-Nearest Neighbours (KNNs) [14] and decision tree (DT) [15]. While these approaches showed promise on controlled datasets, their performance degraded significantly in real-world scenarios where disease symptoms vary in intensity, lighting, and background noise [10]. The reliance on manual feature design also limited the generalisability of these models to diverse disease types in tomato.

Recent advances have seen a shift towards deep learning for end-to-end feature extraction and classification [16]. Several state-of-the-art deep neural networks have been used for recognising diseases from tomato leaves. In [17,18], variants of the ResNet architecture are employed as both feature extractors and classifiers, demonstrating strong performance. To tackle occlusion and small disease regions, Sun et al. [19] proposed a novel approach to boost global-local feature learning, achieving 97.2% mAP50 and outperforming YOLOv10s. In response to the challenges of detecting tomato diseases in complex environments, the TomatoDet model proposed in [20] integrates Swin-DDETR-based self-attention, a Meta-ACON activation function, and an improved BiFPN to enhance small lesion detection, reduce false detections and achieve an mAP of 92.3% with real-time performance. In [21], a YOLOv8-based model achieved 66.67% accuracy for tomato disease detection in Saudi Arabia, which highlights challenges in class differentiation and the need for further improvements. In [22], an improved YOLOv7 model achieved 98.8% accuracy in detecting and classifying seven tomato leaf diseases under field-like conditions.

To facilitate deployment on resource-constrained devices, lightweight models have been developed, offering efficient classification with reduced computational requirements [23]. Le et al. [24] introduced a lightweight hybrid architecture that integrates convolutional and recurrent neural networks for tomato leaf disease classification, demonstrating competitive performance while maintaining computational efficiency. To facilitate real-time tomato leaf disease detection on embedded devices, in [25], an enhanced YOLOv8n model is proposed, which achieved a 64.85% increase in detection speed on Jetson Nano while maintaining high accuracy. Despite the proliferation of lightweight deep learning models for tomato leaf disease recognition, there remains a notable gap in addressing the challenges posed by limited labelled datasets.

2.2. Plant Disease Recognition with Limited Data

Despite the success of deep learning models, their performance often hinges on the availability of large, annotated datasets, which is a major limitation in agricultural contexts where labelled data is scarce. To address this, several studies have explored learning strategies under limited supervision. For instance, in [26], transfer learning with pre-trained CNNs is utilised to compensate for small labelled datasets, while [27] adopts a semi-supervised learning approach using consistency regularisation to improve generalisation with fewer annotations. George et al. [28] conducted a comprehensive comparative analysis of state-of-the-art pre-trained deep neural networks, encompassing both deep and lightweight architectures, for plant disease classification through fine-tuning approaches. Bakır [29] conducted a comprehensive study evaluating the impact of fine-tuning pre-trained CNN architectures on tomato disease detection. The research emphasises the significance of synchronously tuning both the feature extraction and classification phases of CNN models, demonstrating that such an approach can substantially enhance detection performance in complex agricultural environments. Dong et al. [30] investigated the impact of fine-tuning paradigms on out-of-distribution detection for plant diseases. This demonstrates that visual prompt tuning significantly outperforms traditional fine-tuning methods, achieving an AUROC of 94.8% in eight-shot settings, which in turn enhances the recognition of previously unseen plant diseases.

In another recent study, researchers explored the application of Vision Transformer (ViT) models, such as ViT-ImageNet, ViT-Base, and ViT-Small, for tomato leaf disease classification, demonstrating that these transfer learning approaches achieved superior accuracy on both the PlantVillage dataset and the newly introduced TomatoEbola dataset [31]. In a recent study, Alam et al. [32] conducted a comprehensive evaluation of nine widely used pre-trained CNN models, such as DenseNet201, EfficientNetB3, EfficientNetB4, InceptionResNetV2, MobileNetV2, ResNet50, ResNet152, VGG16, and Xception, fine-tuned for plant leaf disease detection. They compared these models against a custom-designed CNN architecture, assessing various performance metrics. The findings revealed that the custom CNN achieved performance comparable to the pre-trained models while demonstrating superior computational efficiency, highlighting its potential for dealing with small labelled data. Thuseethan et al. [2] introduced a novel Siamese-network-based lightweight framework for automatic tomato leaf disease recognition, achieving high accuracy on both the PlantVillage and Taiwan tomato leaf disease datasets. This approach demonstrates effectiveness in handling imbalanced and small datasets with a model comprising approximately 2.96 million trainable parameters.

In the broader context of plant disease recognition, a handful of approaches have been explored to mitigate the limitations posed by scarce labelled data. This includes methods such as few-shot learning, self-supervised learning, and semi-supervised learning to enhance model performance under data constraints. However, these methodologies have not been specifically tailored or extensively applied to tomato leaf disease recognition. Consequently, there is a pressing need for research focused on developing and adapting techniques to effectively address the challenge of limited labelled data within this specific domain.

3. Proposed Method

The proposed TOM-SSL framework addresses the challenge of limited labelled data in tomato leaf disease recognition by leveraging a pseudo-labelling-based SSL approach, which in turn enhances tomato leaf disease recognition performance in scenarios with limited labelled datasets. An overall flow of the TOM-SSL framework is illustrated in Figure 1. Initially, a small fraction (10%) of the available dataset is labelled, while the remaining data is unlabelled. The framework iteratively enhances the training set by incorporating high-confidence pseudo-labelled samples from the unlabelled data, which leads to improved model performance without requiring extensive manual annotation.

3.1. Problem Setting

Let the entire tomato disease dataset be denoted as $\mathcal{D}={\left\{x_i\right\}}_{i=1}^N$, where each $x_i$ represents an image of a tomato leaf. The dataset comprises N samples, each potentially belonging to one of C distinct disease classes, which includes diseased and healthy leaves. To effectively implement the SSL approach, the tomato disease dataset $\mathcal{D}$ is partitioned into two disjoint subsets:

• Labelled Set ${\mathcal{D}}_L={\left\{(x_i,y_i)\right\}}_{i=1}^{N_L}$: This subset contains $N_L$ samples, where each sample $x_i$ is paired with its corresponding ground truth label $y_i\in \{1,2,\dots ,C\}$. In the proposed TOM-SSL framework, $N_L$ is set to 10% of the total tomato disease dataset size, i.e., $N_L=0.1\times N$. This limited labelled set serves as the initial training data for the model.
• Unlabelled Set ${\mathcal{D}}_U={\left\{x_j\right\}}_{j=1}^{N_U}$: Comprising the remaining 90% of the tomato disease dataset, this subset includes $N_U=N-N_L$ samples without associated labels. These unlabelled samples are utilised in the SSL process through pseudo-labelling.

The partitioning strategy employed ensures that the labelled set ${\mathcal{D}}_L$ is representative of all tomato disease classes and the healthy class, maintaining class balance to the extent possible. This is achieved by stratified sampling, where the proportion of each class in ${\mathcal{D}}_L$ mirrors that in the full dataset $\mathcal{D}$. Such a strategy is crucial to prevent class imbalance issues during the initial supervised training phase.

The unlabelled set ${\mathcal{D}}_U$ is then leveraged in subsequent training iterations. The model trained on ${\mathcal{D}}_L$ predicts labels for samples in ${\mathcal{D}}_U$, and high-confidence predictions are incorporated back into the training set as pseudo-labelled data. The threshold of confidence value is set to 0.8 to select the high-confidence samples. This iterative process continues until a stopping criterion is met, which is when no new high-confidence pseudo-labels can be generated.

3.2. Backbone Model

The TOM-SSL framework employs MobileNetV3-Small as its base classifier, denoted as $f_{\theta }$, parameterised by $\theta$. Designed for resource-constrained environments, this architecture is well-suited for tasks like tomato leaf disease recognition, where efficient training with only limited data is crucial. More importantly, the MobileNetV3-Small is optimised for mobile and embedded devices [33]. MobileNetV3-Small was selected as a lightweight backbone suitable for environments with limited labelled data due to its extremely compact architecture and efficient performance. With only  2.5 million parameters and a model size under 10 MB, it delivers strong classification accuracy with very low inference latency, which makes it ideal for SSL on small agricultural datasets [34]. It also combines several advanced techniques such as inverted residual blocks, squeeze-and-excitation (SE) modules, and the hard-swish activation function to achieve a balance between performance and efficiency. Hence, the choice of MobileNetV3-Small ensures that the model remains lightweight and efficient, facilitating deployment in real-world agricultural settings where computational resources may be limited.

Given an input image $x\in {\mathbb{R}}^{H\times W\times 3}$, where H and W denote the height and width, respectively, the model outputs a probability distribution over C disease classes as given below.

$$f_{\theta }\left(x\right)=\mathbf{p}=[p_1,p_2,\dots ,p_C],\mathrm{where}\sum _{c=1}^C{p}_c=1$$

Here, $p_c$ represents the predicted probability that image x belongs to class c.

3.3. Training Procedure

In the semi-supervised TOM-SSL framework, the model initially trains on the labelled dataset ${\mathcal{D}}_L$. Subsequently, it generates pseudo-labels for the unlabelled dataset ${\mathcal{D}}_U$ by predicting labels for unlabelled samples. High-confidence predictions are then incorporated into the training set, effectively expanding ${\mathcal{D}}_L$ with pseudo-labelled data. Here, the confidence threshold is set to 0.80, indicating that a sample is assigned a pseudo-label only if its predicted class probability satisfies $p_i>0.8$ for some $i\in C$, where C denotes the set of all plant disease and healthy classes. The threshold value of 0.8 was determined through a grid search procedure, which systematically evaluated multiple candidate thresholds to identify the value that yielded the best performance on the validation set. This iterative process continues, allowing the model to leverage both labelled and unlabelled data to improve performance.

The training process of the TOM-SSL framework is conducted iteratively, with each iteration consisting of the following key phases. A comprehensive summary of the overall training procedure is presented in Algorithm 1.

Step 1: 

Train the model with labelled data

The model is trained using the cross-entropy loss function, which measures the discrepancy between the predicted class probabilities and the true labels. This loss function encourages the model to assign high probabilities to the correct classes. Given a labelled dataset ${\mathcal{D}}_L={\left\{(x_i,y_i)\right\}}_{i=1}^{N_L}$, where $x_i$ is the i-th input sample and $y_i\in \{1,2,\dots ,C\}$ is the corresponding ground truth label among C classes, the cross-entropy loss is defined as

$${\mathcal{L}}_{\mathrm{CE}}=-{\displaystyle \frac{1}{N_L}}\sum _{i=1}^{N_L}\sum _{c=1}^C{\mathbb{I}}_{[y_i=c]}log{p}_{i,c}$$

where $p_{i,c}$ denotes the predicted probability of class c for the i-th sample, typically obtained using a softmax activation over the model’s output logits. ${\mathbb{I}}_{[y_i=c]}$ is an indicator function that equals 1 if the true label $y_i$ is equal to c, and 0 otherwise. This ensures that only the log-probability corresponding to the correct class contributes to the loss.

The term is averaged over all labelled samples $N_L$ to normalise the loss and provide consistent gradient updates during backpropagation. By minimising ${\mathcal{L}}_{\mathrm{CE}}$, the model is optimised to predict probability distributions that align closely with the true label distribution of the training data.

Step 2: 

Predict pseudo-labels for the unlabelled data

After training on ${\mathcal{D}}_L$, the model predicts labels for the unlabelled data ${\mathcal{D}}_U$. For each unlabelled sample $x_j\in {\mathcal{D}}_U$, the model computes the predicted probabilities ${\mathbf{p}}^{\left(j\right)}=f_{\theta }\left(x_j\right)$. A pseudo-label ${\widehat{y}}_j$ is assigned if the maximum predicted probability exceeds a predefined confidence threshold $\tau$:

$${\widehat{y}}_j=arg\underset{c}{max}{p}_c^{\left(j\right)}\mathrm{if}\underset{c}{max}{p}_c^{\left(j\right)}\ge \tau$$

The set of pseudo-labelled samples is denoted as

$${\mathcal{D}}_P=\{(x_j,{\widehat{y}}_j)\mid x_j\in {\mathcal{D}}_U,\underset{c}{max}{p}_c^{\left(j\right)}\ge \tau \}$$

Step 3: 

Select high confidence samples

In SSL, to leverage additional training signals from unlabelled data, the pseudo-labelled set ${\mathcal{D}}_P$, which consists of confidently predicted labels ${\widehat{y}}_j$ for unlabelled samples $x_j$, is incorporated into the labelled dataset. This results in an augmented labelled dataset that includes both the original labelled samples and the newly pseudo-labelled ones:

$${\mathcal{D}}_L\leftarrow {\mathcal{D}}_L\cup {\mathcal{D}}_P$$

This pseudo-label augmentation enriches the supervision available to the model during training. In parallel, to prevent duplicate or inconsistent learning from the same data points, the corresponding pseudo-labelled samples are removed from the unlabelled dataset ${\mathcal{D}}_U$:

$${\mathcal{D}}_U\leftarrow {\mathcal{D}}_U\setminus \{x_j\mid (x_j,{\widehat{y}}_j)\in {\mathcal{D}}_P\}$$

This ensures that each sample is used either as an unlabelled instance or as a pseudo-labelled one during training, but not both simultaneously. By iteratively incorporating high-confidence pseudo-labels into the labelled set and pruning them from the unlabelled pool, the model progressively refines its performance using a semi-supervised learning strategy.

Step 4: 

Retraining with the pseudo and labelled data

The overall training process proceeds iteratively, alternating between supervised learning on the labelled data and the generation of pseudo-labels for the most confident unlabelled samples. In each iteration, the model is retrained on the augmented labelled dataset ${\mathcal{D}}_L$, and a new set of high-confidence pseudo-labels is generated from the current unlabelled dataset ${\mathcal{D}}_U$. These pseudo-labelled samples are then incorporated into ${\mathcal{D}}_L$ while being removed from ${\mathcal{D}}_U$, as described earlier.

This process is repeated until one of the following stopping criteria is satisfied:

• The unlabelled dataset ${\mathcal{D}}_U$ becomes empty, meaning that all available samples have been confidently assigned pseudo-labels and added to the labelled set.
• The number of pseudo-labelled samples selected in the current iteration falls below a predefined threshold, indicating that the model is no longer able to assign labels to additional unlabelled samples confidently. This serves as a convergence criterion to avoid overfitting or the propagation of incorrect labels.

By adopting this iterative approach, the model incrementally improves its performance by utilising both the ground-truth labels and the high-confidence predictions, thereby achieving a more effective semi-supervised learning framework.

The supervised learning and pseudo-labelling steps are repeated iteratively until one of the following stopping criteria is met:

• The unlabelled set ${\mathcal{D}}_U$ becomes empty.
• The number of pseudo-labelled samples in an iteration falls below a predefined threshold, indicating convergence.

Algorithm 1 TOM-SSL Training Procedure

Require: Labelled dataset ${\mathcal{D}}_L$, unlabelled dataset ${\mathcal{D}}_U$, confidence threshold $\tau$

Ensure: Trained model $f_{\theta }$

1: while ${\mathcal{D}}_U$ is not empty do

2:    Train $f_{\theta }$ on ${\mathcal{D}}_L$ using cross-entropy loss

3:    Initialise ${\mathcal{D}}_P\leftarrow \mathsf{\varnothing }$

4:    for each $x_j\in {\mathcal{D}}_U$ do

5:        Compute ${\mathbf{p}}^{\left(j\right)}=f_{\theta }\left(x_j\right)$

6:        if ${max}_c{p}_c^{\left(j\right)}\ge \tau$ then

7:           Assign ${\widehat{y}}_j=arg{max}_c{p}_c^{\left(j\right)}$

8:           ${\mathcal{D}}_P\leftarrow {\mathcal{D}}_P\cup \left\{(x_j,{\widehat{y}}_j)\right\}$

9:        end if

10:    end for

11:    ${\mathcal{D}}_L\leftarrow {\mathcal{D}}_L\cup {\mathcal{D}}_P$

12:    ${\mathcal{D}}_U\leftarrow {\mathcal{D}}_U\setminus \{x_j\mid (x_j,{\widehat{y}}_j)\in {\mathcal{D}}_P\}$

13:    if size of ${\mathcal{D}}_P$ < threshold then

14:        break

15:    end if

16: end while

17: return  $f_{\theta }$

Upon completion of the iterative training process, the final model $f_{\theta }$ is evaluated on a separate test set ${\mathcal{D}}_{\mathrm{test}}={\left\{(x_k,y_k)\right\}}_{k=1}^{N_{\mathrm{test}}}$ using standard classification metrics. Next, the experiments and results are discussed.

4. Experiments and Results

This section presents a detailed description of the dataset, outlines the experimental setup, and describes the experiments conducted. Furthermore, the results are reported and discussed.

4.1. Datasets

The proposed TOM-SSL framework for tomato leaf disease recognition is evaluated on two publicly available benchmark datasets (https://data.mendeley.com/datasets/ngdgg79rzb/1, accessed on 1 July 2025). One is a subset of the PlantVillage dataset comprising tomato-specific disease samples, and the other is the Taiwan Tomato Disease dataset. Figure 2 presents one representative image sample from each class in both datasets.

The tomato subset of the PlantVillage dataset comprises 14,461 images categorised into ten distinct classes, including nine disease types and one healthy class. The disease classes included are Early Blight (800), Bacterial Spot (1702 images), Leaf Mould (762), Late Blight (1528), Target Spot (1124), Septoria Leaf Spot (1417), Mosaic Virus (299), Two-spotted Spider Mite (1341) and Yellow Leaf Curl Virus (4286). The rest of the samples are for the Healthy class. This dataset is notably imbalanced (illustrated in Figure 3 (left)), with a significant disparity in class distribution; the tomato yellow leaf curl virus class, as the majority, contains nearly fourteen times more images than the tomato mosaic virus class, which has the fewest samples. All images depict a single tomato leaf on a uniform background and have been resized to 227 × 227 pixels for consistency in model input.

The original Taiwan Tomato Disease Dataset is relatively limited in size and is utilised to assess the performance of the proposed model. It comprises 622 tomato leaf images, each resized to 227 × 227 pixels and distributed across six categories, including both healthy and diseased classes: Black Mould (67), Bacterial Spot (100 samples), Grey Spot (84), Late Blight (98), Powdery Mildew (157) and Healthy (106). To address the limited sample size, an augmented version of the dataset is created by applying a range of image augmentation techniques. These include rotations of 90°, 180° and 270°; horizontal and vertical mirroring; and brightness adjustments. This dataset exhibits a relatively lower degree of class imbalance (illustrated in Figure 3 right) compared to the tomato subset of the PlantVillage dataset. The resulting augmented dataset significantly increases the number of samples, yielding a total of 4976 tomato leaf images.

4.2. Experimental Settings

4.2.1. Environments

A lightweight MobileNetV3-Small model, pretrained on the ImageNet dataset, serves as the backbone of the TOM-SSL framework. Distinct training configurations are employed for the initialisation and retraining phases of the model. During the initialisation phase, the model is trained for 100 epochs using stochastic gradient descent (SGD) with a mini-batch size of 32 and a momentum value of 0.9. The initial learning rate is set to 0.01 and follows a step-down decay policy, with a decay rate of 0.001 (i.e., 1/1000). In contrast, each retraining phase consists of 50 training epochs, with a reduced mini-batch size of 16. The initial learning rate and decay rate for this phase are set to 0.001 and 0.0005, respectively, while maintaining the same SGD optimiser configuration with a momentum of 0.9. To mitigate overfitting, L2 regularisation is applied throughout both phases. Furthermore, early stopping is utilised in both training stages to prevent unnecessary iterations once performance ceases to improve.

4.2.2. Implementations and Evaluation Metrics

The proposed framework, along with the baseline methods for comparison, is implemented using the open-source PyTorch machine learning library (https://pytorch.org/, accessed on 16 June 2025). Model training is conducted within the Google Colab environment (https://colab.research.google.com/, accessed on 16 June 2025). Performance evaluation is conducted on a system equipped with an Intel Xeon CPU @ 2.30 GHz, supported by an NVIDIA Tesla T4 GPU with 13 GB of RAM.

To evaluate the performance of the proposed method, two widely adopted metrics, namely the average accuracy and macro F1 score, are employed. Accuracy provides a direct measure of how closely the predicted tomato disease class matches the actual class label. However, in the presence of class imbalance, accuracy alone may not be sufficient to capture model performance across all categories. The macro F1 score addresses this limitation by computing the unweighted mean of the F1 scores for each class, thereby giving equal importance to all classes regardless of their frequency in the dataset. This combination offers a balanced and interpretable assessment for imbalanced multi-class classification. Additional metrics were omitted, as macro F1-score already incorporates both precision and recall effectively. The accuracy and macro F1 score are defined as

$$\mathrm{Accuracy}={\displaystyle \frac{tp+tn}{tp+tn+fp+fn}},\mathrm{Macro}\mathrm{F}1={\displaystyle \frac{1}{N}}\sum _{i=1}^N{\displaystyle \frac{2·tp}{2·tp+fp+fn}}$$

(1)

where $tp$, $tn$, $fp$, and $fn$ represent the number of true positives, true negatives, false positives, and false negatives, respectively.

4.3. Results and Discussion

This section presents and analyses the performance of the proposed TOM-SSL framework on the PlantVillage and Taiwan tomato disease datasets. The results are reported in terms of average accuracy and macro F1 score and are compared against fully supervised and other semi-supervised baselines.

4.4. Performance on the PlantVillage Dataset

Fully supervised approaches have consistently achieved very good performance in tomato leaf disease classification and are widely regarded as the state of the art in image-based plant disease recognition [35]. Supervised deep learning methods remain the de facto benchmark, while alternative learning paradigms, particularly those not leveraging deep hierarchical feature representations, demonstrate substantially lower accuracy in tomato leaf disease recognition frameworks [36]. Hence, only the supervised techniques and alternative semi-supervised techniques are compared here against the proposed approach.

Table 1. Performance comparison on the PlantVillage tomato subset using 10% labelled data. The highest accuracy and Macro F1 scores are highlighted in bold.

Category	Model	Accuracy (%)	Macro F1 (%)
Fully Supervised	ViT-Tiny	58.78	53.63
	MobileViT-Small	55.31	53.17
	EfficientNet-B0	45.71	42.92
	EfficientViT-B2	49.17	48.80
	Swin-Tiny	47.31	44.11
	ConvNeXt-Tiny	40.17	37.82
	ConvNeXt-Small	45.51	40.71
	MobileNetV3-Small	59.39	54.17
Semi-Supervised	Mean Teacher	61.74	59.27
	VAT	62.18	58.78
	SimSiam	67.31	61.10
	MoCo (with fine-tuning)	71.41	62.25
	MixMatch	70.11	61.51
	TOM-SSL	72.51	65.46

presents a comparative evaluation of multiple fully supervised and semi-supervised models on the PlantVillage tomato subset, conducted using only 10% of the labelled data. Among the fully supervised approaches, MobileNetV3-Small and Vision Transformer Tiny (ViT-Tiny) achieve the highest accuracy (59.39% and 58.78%, respectively) and macro F1 scores (54.17% and 53.63%, respectively), outperforming other architectures such as the EfficientNet-B0, Swin-Tiny, and ConvNeXt variants. These findings indicate that lightweight transformer-based models and compact convolutional networks can achieve comparatively strong performance under limited labelled data conditions. However, their accuracy remains insufficient for dependable disease classification in real-world applications. Notably, the performance of ConvNeXt-Tiny is the lowest in both metrics, highlighting its sensitivity to label scarcity.

In contrast, the semi-supervised methods demonstrate a clear performance advantage over their fully supervised counterparts. Classical consistency-regularisation approaches such as VAT and Mean Teacher surpass most supervised baselines, with VAT achieving 62.18% accuracy and 58.78% macro F1. More advanced contrastive learning methods like SimSiam, MoCo, and MixMatch continue this upward trend, with MoCo achieving 71.41% accuracy and 62.25% macro F1. More importantly, as can be seen, the proposed TOM-SSL framework further improves upon these results, achieving the highest scores in both accuracy (72.51%) and macro F1 (65.46%). This superior performance highlights the effectiveness of TOM-SSL in leveraging both labelled and unlabelled data through an efficient pseudo-labelling strategy.

The confusion matrix presented in Figure 4 reveals that the TOM-SSL framework performs strongly on several classes, particularly Yellow Leaf Curl and Bacterial Spot, with class-specific accuracies of 86% and 74.9%. Similarly, the classes Healthy and Late Blight exhibit relatively strong performance, whereas others, such as Mosaic Virus, Leaf Mould and Early Blight, achieve class-specific accuracies below 60%. There is notable confusion between the Yellow Leaf Curl and Mosaic Virus classes, with 8.0% of Yellow Leaf Curl samples misclassified as Mosaic Virus and 17.9% of Mosaic Virus samples incorrectly predicted as Yellow Leaf Curl. This is mainly due to the biological similarities in how symptoms appear, which can cause overlap between different conditions and make them harder to distinguish accurately. Similarly, substantial misclassification is observed between Septoria Leaf Spot and Bacterial Spot, as well as between Late Blight and Early Blight, indicating that these class pairs exhibit overlapping visual characteristics that challenge the model’s discriminative capability.

4.5. Performance on the Taiwan Tomato Disease Dataset

Table 2. Performance comparison on the Taiwan tomato dataset using 10% labelled data. The highest accuracy and Macro F1 scores are highlighted in bold.

Category	Model	Accuracy (%)	Macro F1 (%)
Fully Supervised	ViT-Tiny	57.63	54.25
	MobileViT-Small	51.87	48.39
	EfficientNet-B0	42.37	39.75
	EfficientViT-B2	45.39	43.15
	Swin-Tiny	44.71	41.23
	ConvNeXt-Tiny	37.03	34.18
	ConvNeXt-Small	42.15	39.87
	MobileNetV3-Small	56.52	53.73
Semi-Supervised	Mean Teacher	59.53	56.71
	VAT	60.37	58.63
	SimSiam	66.28	64.69
	MoCo (with fine-tuning)	69.63	67.14
	MixMatch	67.26	65.43
	TOM-SSL	70.87	69.28

presents a comparative evaluation of various deep learning models under both fully supervised and semi-supervised learning settings on the Taiwan tomato dataset, using only 10% of the labelled data. The fully supervised models, despite being widely adopted in vision tasks, show limited performance under label-scarce conditions. For instance, ViT-Tiny and MobileNetV3-Small yield the highest accuracy scores within this category (57.63% and 56.52%, respectively), while ConvNeXt-Tiny lags significantly behind with only 37.03% accuracy and 34.18% Macro F1. These results suggest that relying solely on limited annotated data restricts the learning capacity of even state-of-the-art architectures.

Conversely, the semi-supervised models demonstrate clear performance improvements across all metrics. Classical methods like Mean Teacher and VAT deliver noticeable gains over the supervised baselines, while modern contrastive and consistency-based methods such as SimSiam, MoCo and MixMatch exhibit substantial boosts in both accuracy and Macro F1. Notably, the proposed TOM-SSL framework achieves the best performance, reaching 70.87% accuracy and 69.28% macro F1. This reinforces the value of semi-supervised learning in low-resource contexts and suggests that TOM-SSL’s architecture, likely benefiting from an integrated pseudo-labelling strategy, is particularly effective in extracting discriminative features from unlabelled data.

The confusion matrix illustrated in Figure 5 depicts the class-wise classification performance across six categories in the Taiwan dataset. Powdery Mildew demonstrates relatively high classification performance, with 78.1% of its samples correctly identified, although a few were misclassified as Late Blight and Bacterial Spot. In contrast, the Black Mould class demonstrates significant misclassification, with only 50% of the samples correctly identified, while a considerable proportion are erroneously classified as Bacterial Spot and Grey Spot. The Healthy, Bacterial Spot, and Late Blight classes exhibit moderate classification performance, with accuracies of 77.3%, 72.7% and 70%, respectively. Frequent misclassifications are observed between the Black Mould and Grey Spot classes, as well as between the Late Blight and Powdery Mildew classes.

4.6. Discussion

The experimental findings across both the PlantVillage and Taiwan tomato disease datasets consistently demonstrate the effectiveness of the proposed TOM-SSL framework in semi-supervised settings with limited annotated data. In both datasets, TOM-SSL achieves the highest accuracy and macro F1 scores compared to existing fully supervised and semi-supervised baselines. These results highlight the model’s ability to effectively exploit unlabelled data through its pseudo-labelling strategy, which enhances generalisation performance and reduces dependence on extensive labelled datasets. This capability is particularly beneficial in agricultural disease detection, where manual annotation is both labour-intensive and resource-demanding.

A detailed analysis of the confusion matrices reveals the challenges and strengths of the model in class-wise prediction. TOM-SSL demonstrates high discriminative capability for certain disease classes such as Yellow Leaf Curl, Powdery Mildew and Healthy, suggesting that the model effectively captures distinctive visual patterns for these categories. However, consistent misclassifications between visually similar disease pairs—such as Yellow Leaf Curl vs. Mosaic Virus, Black Mould vs. Grey Spot, and Late Blight vs. Early Blight—highlight the limitations posed by subtle inter-class variations and noisy or overlapping image features. These confusions suggest potential areas for enhancement, such as incorporating attention mechanisms, domain-specific augmentations, or multi-modal data (e.g., spectral or temporal features) to improve the model’s discriminative power. Overall, TOM-SSL proves to be a promising framework for practical deployment in low-label agricultural contexts while also leaving room for future improvements in fine-grained disease recognition.

5. Conclusions

Labelled data is often limited in the agricultural domain due to various factors, such as the high cost of expert annotation, seasonal availability of crop conditions and the diverse and complex nature of plant diseases. These challenges significantly hinder the scalability and generalisation of supervised learning approaches in real-world agricultural applications. In particular, expert labelling often requires domain-specific knowledge that is not readily accessible in many agricultural regions, while variations in plant phenotypes across seasons and geographical locations further complicate the annotation process. Additionally, the visual similarities between different disease symptoms and between healthy and early-stage diseased plants reduce the reliability of manual labelling, leading to inconsistencies in training datasets. As a result, models trained purely on limited labelled data often suffer from poor generalisation when deployed in unseen or dynamic environments. This limitation calls for alternative learning paradigms that can effectively utilise the vast amount of unlabelled agricultural data while reducing reliance on costly annotations.

The proposed TOM-SSL framework demonstrated strong potential in addressing the challenges of tomato disease classification under limited labelled data conditions by leveraging the power of semi-supervised learning. Through extensive evaluations on both the PlantVillage and Taiwan tomato disease datasets, TOM-SSL consistently outperforms existing fully supervised and semi-supervised baselines, achieving the highest accuracy and macro F1 scores using only 10% of labelled data. These improvements can be attributed to its robust pseudo-labelling mechanism and tailored design, which effectively utilise unlabelled data to enhance generalisation capabilities. The confusion matrix analysis further reveals that TOM-SSL maintains strong class-wise prediction performance across several disease types, although confusion persists among certain visually similar classes, which indicates opportunities for refinement.

As future work, the TOM-SSL framework can be extended to include domain-specific data augmentation strategies or attention mechanisms to improve class separability. These strategies may incorporate prior knowledge of plant morphology to generate more realistic variations, which in turn enhance the model’s ability to distinguish between visually similar disease classes. Furthermore, integrating TOM-SSL with lightweight architectures could facilitate deployment on edge devices, which enables real-time disease diagnosis in resource-constrained settings such as rural farms. Evaluating the model’s robustness under challenging environmental conditions, such as varying lighting, occlusion, motion blur, and field-level noise, is also essential to ensure its reliability in practical agricultural scenarios. In addition, future research may explore the combination of TOM-SSL with active learning or continual learning paradigms, allowing the model to selectively query informative samples and adapt incrementally to new disease types or domain shifts.