Tomato Disease Classification using Transformer- Based Deep Learning Models 1 st Xiaocun Huang Faculty of Information Technology City University Malaysia Cyberjaya, Malaysia huangxiaocun@caztc.edu.cn 3 rd Mohd Nurul Hafiz Ibrahim Faculty of Information Technology City University Malaysia Cyberjaya, Malaysia mohd.nurul.hafiz@city.edu.my 2 nd Mustafa Muwafak Alobaedy Faculty of Information Technology City University Malaysia Cyberjaya, Malaysia alobaedy@ieee.org 4 th Ali A Khalaf Computer Science Department, College of Science University of Baghdad Baghdad, Iraq alrahmanali@sc.uobaghdad.edu.iq Abstract— Crop diseases present a major challenge to global agriculture, reducing yields and threatening food security. Traditional manual methods for detecting diseases in crops like tomatoes are labor-intensive and lack accuracy. This paper proposes a deep learning model based on the Transformer architecture for automated tomato leaf disease classification. Leveraging the self-attention mechanism of Transformers, the model captures global contextual features from leaf images, achieving superior recognition accuracy. The model was trained and tested on a diverse dataset comprising ten categories (nine diseases and healthy leaves), with data augmentation enhancing its robustness. Experimental results demonstrate that the Transformer-based model achieved a classification accuracy of 91.77%, outperforming traditional CNN models such as VGG19 and ResNet50. This study highlights the potential of Transformer architectures in precision agriculture, enabling efficient and accurate disease recognition. Future work will focus on optimizing the model for deployment of mobile devices to facilitate real-time, on-field disease diagnosis. Keywords— Transformer model, tomato disease classification, deep learning, attention mechanism, computer vision. I. INTRODUCTION Plant diseases significantly affect crop yields and food safety, directly impacting the agricultural economy and causing substantial economic losses [1]. Moreover, they exacerbate the global issue of food insecurity [2]. To mitigate losses caused by plant diseases, it is essential to continuously monitor crop health and implement timely and effective countermeasures [3]. However, traditional manual detection methods are unsuitable for large-scale monitoring, as they require substantial human and financial resources. Additionally, these methods often fail to detect disease progression promptly, leading to missed opportunities for intervention and, consequently, increased losses [4]. Therefore, accurate and timely detection of plant diseases is critical. In recent years, the rapid development of artificial intelligence (AI) has enabled many fields to adopt AI-based technologies to address complex challenges effectively [5–7]. As a crucial branch of AI, computer vision has emerged as a powerful tool for solving various visual recognition problems, contributing significantly to advances in object detection and classification [8]. Convolutional Neural Networks (CNNs), in particular, have demonstrated outstanding performance in image classification tasks. Notable CNN-based architectures include EfficientNet [9], AlexNet [10], ResNet [11], DenseNet [12], and MobileNet [13]. Despite their success, CNNs have inherent limitations, especially regarding their restricted receptive field, which limits their ability to capture global image context due to the fixed size of convolutional kernels. To address this limitation, researchers have drawn inspiration from the Transformer architecture, initially developed for natural language processing tasks [14], and adapted it for computer vision applications through the Vision Transformer (ViT) [15]. ViT has shown promising results in various computer vision tasks such as image classification [16], object detection [17,18], and image segmentation [19]. Owing to the self-attention mechanism, Vision Transformers can effectively capture global dependencies, allowing the model to process holistic image information from the beginning of training. This study leverages the Transformer architecture to develop a deep learning model for classifying tomato leaf diseases. The proposed model demonstrates high recognition accuracy, highlighting the potential of Transformer-based models in agricultural disease detection and precision farming. II. CONVOLUTIONAL NEURAL NETWORKS The traditional Convolutional Neural Network (CNN) architecture consists of five main components: the input layer, convolutional layer, pooling (down sampling) layer, activation function layer, and fully connected layer. As illustrated in Figure 1, these components are sequentially stacked to form a deep convolutional structure, enabling the network to effectively extract features from input images. Fig. 1. Architecture of a Convolutional Neural Network (CNN) [20]. The size of the convolutional kernel determines the receptive field and feature extraction capability of a neural network model. A standard convolutional layer can include kernels of various sizes, such as 3×3, 5×5, and 7×7, each affecting the spatial resolution and receptive field of the resulting feature maps. The pooling layer plays a critical role in convolutional neural networks by performing down sampling operations. It not only helps extract essential features but also reduces the computational cost and the number of parameters in the model. Typically, the pooling layer follows the convolutional layer and processes its output feature maps. Activation functions are another essential component, enabling the network to learn complex patterns in image data. By introducing non-linearity into the model, activation functions allow neural networks to represent and approximate intricate functional relationships, thereby improving their ability to fit complex data. The fully connected layer, usually positioned at the end of a convolutional neural network, serves to transform the twodimensional feature maps into one-dimensional vectors. It integrates the local features extracted by previous layers into a comprehensive representation of the input image using a weight matrix. III. TRANSFORMER MODEL STRUCTURE In 2017, Vaswani et al. proposed the Transformer model for natural language processing tasks. Inspired by this architecture, Dosovitskiy et al. later adapted the Transformer for computer vision applications. Since then, the Transformer has demonstrated outstanding performance in various core computer vision tasks. The Transformer model consists of two main components: the encoder and the decoder. In most vision-related applications, the encoder plays the primary role, focusing on feature extraction and representation. The encoder is composed of multiple identical layers, known as Encoder Layers, each containing two key subcomponents: a Self- Attention layer and a Feed-Forward layer. This structure is illustrated in Figure 2. In the fine-grained classification task of tomato leaf diseases, the Transformer model performs feature encoding by dividing the plant image into patches. These image patches are then flattened into a sequential format. The model's selfattention mechanism assigns different weights to each element in the sequence, enabling it to effectively capture the relationships between patches. This process enhances the model’s ability to recognize subtle variations in the leaf images, as illustrated in Figure 3. Fig. 2. Transformer Model Architecture Fig. 3. Feature extraction of tomato leaf diseases.[21] IV. DATA The dataset used in this study comprises images from ten categories, including various tomato leaf diseases and healthy leaves. The images were collected from multiple sources, with some obtained directly and others sourced from the internet, resulting in variations in format and resolution. To evaluate the model’s performance, the dataset was divided into training, validation, and testing sets in a 60:20:20 ratio. Data augmentation—a commonly used technique in deep learning—was applied to increase the dataset size and diversity by transforming and expanding the existing images. This approach helps improve the model’s generalization and robustness, enabling it to make more accurate predictions on unseen data. During preprocessing, techniques such as flipping, mirroring, and noise addition were used to augment the images. The distribution of images across the ten classes, including diseased and healthy tomato leaves, is presented in Table I. TABLE I. DATASET DISTRIBUTION disease species Number Bacterial_spot 1735 Early_blight 2236 Late_blight 2053 Leaf_Mold 2175 Septoria_leaf_spot 2349 Spider_mites Two-spotted_spider_mite 1744 Target_Spot 1826 Tomato_mosaic_virus 1818 Tomato_Yellow_Leaf_Curl_Virus 1989 Healthy 1952 V. EXPERIMENT AND ANALYSIS The experimental environment used for model training and evaluation is outlined in Table II, detailing both the hardware and software specifications. These configurations were selected to ensure smooth processing of image data and efficient training of the Transformer-based model. TABLE II. HARDWARE AND SOFTWARE REQUIREMENTS The input images were resized to 224×224 pixels and then divided by the Transformer model into 16×16 patches, each of size 14×14. The model was trained using the cross-entropy loss function and optimized with Stochastic Gradient Descent (SGD). Training was conducted for 20 epochs with a batch size of 8. The initial learning rate was set to 0.001 and dynamically adjusted using a cosine decay schedule. The model achieved an accuracy of 91.77%, precision of 91.89%, recall of 91.77%, and an F1-score of 91.77% across 10 categories, comprising nine disease types and one healthy class. Figures 4 and 5 illustrate the training accuracy and loss curves, respectively. Fig. 4. Training and validation accuracy curve of the Transformer model over 20 epochs. Fig. 5. Training and validation loss curve of the Transformer model over 20 epochs. As illustrated in the confusion matrix (Figure 6), the model demonstrates high recognition rates for six classes—Bacterial Spot, Leaf Mold, Septoria Leaf Spot, Spider Mites (Two- Spotted Spider Mite), Tomato Yellow Leaf Curl Virus, and Tomato Mosaic Virus—all exceeding 90%. Recognition rates for Early Blight, Healthy, and Late Blight are slightly lower but still approach 90%. In contrast, the Target Spot class records the lowest recognition rate. This is likely due to the relatively small and faint spots on the leaves, which become less visible under light exposure, causing the model to misclassify them as healthy. To compare the proposed Transformer model with conventional CNN architecture, we conducted additional experiments using VGG19 and ResNet50. The results demonstrate that the Transformer model outperforms both CNN models in terms of accuracy, precision, recall, and F1score. As illustrated in Figures 7–10, classification metrics were calculated separately for each model. Specifically, VGG19 achieved an accuracy of 84.23%, ResNet50 achieved 86.75%, while the Transformer reached 91.77%. In terms of precision, VGG19 recorded 85.71%, ResNet50 87.62%, and the Transformer achieved 91.89%. The recall for the Transformer increased by approximately 7.54% over VGG19 and 5.02% over ResNet50. Regarding F1score, VGG19 achieved 84.23%, ResNet50 86.75%, and Transformer again led with 91.77%. Hardware Software 8GB RAM Microsoft Windows 2010 Core i5-11320H Processor Python 3.8 500GB of Hard drive capacity Anaconda 3 GeForce RTX 3060 TensorFlow 2.0 Fig. 6. Normalized confusion matrix of the Transformer model. Fig. 7. Comparison of classification accuracy across VGG19, ResNet50, and Transformer models. Fig. 8. Comparison of precision scores for VGG19, ResNet50, and Transformer models. Fig. 9. F1-score comparison among VGG19, ResNet50, and Transformer models. Fig. 10. Recall comparison across VGG19, ResNet50, and Transformer models. VI. CONCLUSION This study demonstrates the effectiveness of using a Transformer-based deep learning model for the classification and recognition of tomato leaf diseases. The self-attention mechanism within the Transformer architecture enables the model to focus selectively on critical regions of input images, amplifying relevant features while suppressing noise and irrelevant details. As a result, the model achieves high classification performance, with an accuracy of 91.77%, outperforming traditional CNN-based approaches. However, due to the model’s complexity and computational requirements, it is currently not optimized for deployment on resource-constrained devices such as smartphones or edge devices. Future work will focus on optimizing and compressing the model to enable real-time, mobile-based disease detection. This advancement could significantly enhance precision agriculture by providing accessible, rapid, and reliable diagnostic tools to farmers and agricultural workers in the field. R EFERENCES [1] J. J. Burdon, L. G. Barrett, L N. Yang, et al. Maximising world food production through disease control[J]. BioScience, 2020, 70(2): 126- 128. [2] Z. Li, R. Paul, T. Ba Tis, et al. Non-invasive plant disease diagnostics enabled by smartphone-based fingerprinting of leaf volatiles[J]. Nature Plants, 2019, 5(8): 856866. [3] M. A. Altieri. Agroecology: the science of sustainable agriculture[M]. CRC Press, 2018. [4] S. P. Mohanty, D. P. Hughes, M. Salathé. Using deep learning for image-based plant disease detection[J]. Frontiers in plant science, 2016, 7: 1419. [5] G. Litjens, T. Kooi, B. E. Bejnordi, et al. A survey on deep learning in medical image analysis[J]. Medical image analysis, 2017, 42: 6088. [6] W. Min, S. Jiang, L. Liu, et al. A survey on food computing[J]. ACM Computing Surveys (CSUR), 2019, 52(5): 136. [7] E. Moen, D. Bannon, T. Kudo, et al. Deep learning for cellular image analysis[J]. Nature Methods, 2019, 16(12): 12331246. [8] J. G. Arnal Barbedo. Digital image processing techniques for detecting, quantifying, and classifying plant diseases[J]. SpringerPlus, 2013, 2(1): 112. [9] M. Tan, Q. Le. Efficientnet: Rethinking model scaling for convolutional neural networks[C]//International conference on machine learning. 2019: 61056114. [10] A. Krizhevsky, I. Sutskever, G. E. Hinton. Imagenet classification with deep convolutional neural networks[J]. Advances in neural information processing systems, 2012, 25: 10971105. [11] K. He, X. Zhang, S. Ren, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 770778. [12] G. Huang, Z. Liu, L. Van Der Maaten, et al. Densely connected convolutional networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 47004708. [13] A. G. Howard, M. Zhu, B. Chen, et al. Mobilenets: Efficient convolutional neural networks for mobile vision applications[J]. ArXiv preprint arXiv:1704.04861, 2017. [14] A. Vaswani, N. Shazeer, N. Parmar, et al. Attention is all you need[J]— advances in neural information processing systems, 2017, 30. [15] A. Dosovitskiy, L. Beyer, A. Kolesnikov, et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale[C]//International Conference on Learning Representations. 2020. [16] H. Touvron, M. Cord, M. Douze, et al. Training data efficient image transformers&distillation through attention[C]//International Conference on Machine Learning. 2021: 1034710357. [17] N. Carion, F. Massa, G. Synnaeve, et al. End-to-end object detection with transformers[C]//European Conference on Computer Vision. 2020: 213229. [18] X. Zhu, W. Su, L. Lu, et al. Deformable detr: Deformable transformers for end-to-end object detection[J]. ArXiv preprint arXiv:2010.04159, 2020. [19] L. Ye, M. Rochan, Z. Liu, et al. Cross-modal self-attention network for referring image segmentation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019: 1050210511. [20] Zhang S, Zhang Q, Li H. A review of sign language recognition based on deep learning [J]. Journal of Electronics and Information Technology, 2020, 42 (4): 1021-1032. GX. Hao. Research on tomato leaf disease recognition based on deep learning[D]. Tarim University, 2023.