HybridViT for robust wheat leaf disease detection using CLAHE and attention-based feature fusion

Abstract

Wheat (Triticum aestivum L.) is a staple crop of paramount importance to global food security; however, its productivity is significantly compromised by foliar diseases. Conventional diagnostic approaches, relying on manual observation or laboratory analyses, are often labor-intensive and susceptible to inaccuracies. While recent advancements in deep learning present promising avenues for automated disease detection, persistent challenges such as limited annotated datasets, environmental heterogeneity, and model generalization continue to hinder optimal performance. This study proposes a novel hybrid deep learning model called HybridViT, which combines ConvNeXt and Vision Transformer (ViT) architectures with the Convolutional Block Attention Module (CBAM) to improve the classification of wheat leaf diseases. While ConvNeXt ensures local feature extraction and ViT provides global contextual understanding, CBAM dynamically highlights the most discriminative features. Additionally, the Contrast Limited Adaptive Histogram Equalization (CLAHE) method is employed to enhance the visibility of disease symptoms in low-contrast leaf images. Unlike conventional hybrid CNN–Transformer approaches that rely on static feature concatenation, the proposed model employs an adaptive gated fusion mechanism to dynamically balance local and global feature representations. The fused features are further refined using a lightweight CBAM module to enhance discriminative capability. Additionally, Contrast Limited Adaptive Histogram Equalization (CLAHE) is applied to improve feature visibility under varying illumination conditions. Evaluated on three different datasets obtained under both controlled and field conditions, HybridViT achieved 100% accuracy on balanced datasets and 99.10% accuracy on complex images captured in real-world conditions, surpassing existing methods. Furthermore, a 5-fold cross-validation strategy yielded an average accuracy of 99.04% ± 0.22, demonstrating the model’s robustness and stability across different data splits. The results demonstrate the model’s robustness against environmental noise, lighting variations, and class imbalance. This approach, which enables early and accurate disease diagnosis, supports sustainable agricultural practices, reduces pesticide use, and contributes to global food security.

Introduction

Wheat (Triticum aestivum L.) is a strategic cereal crop that serves as a primary source of minerals, vitamins, protein, and energy for a large portion of the global population. It is the main raw material used in the production of flour, which is essential for manufacturing bread, pasta, biscuits, and various other food products¹. Contributing approximately 20% of the global caloric intake, wheat is the third most consumed staple food after maize and rice^2–4. Among agricultural commodities, wheat has consistently held an indispensable place in global nutrition due to its high nutritional value and excellent storage properties, making it one of the cornerstones of modern food production. Especially in developing countries such as Türkiye, wheat is considered a vital food source and is recognized as a strategic crop⁵. As a result of this strategic status, wheat is not only significant at the national level but also critical to global food security and economic stability. In regions such as Asia, Africa, and the Middle East, it constitutes the foundation of agricultural production and supports the sustenance and livelihoods of millions of people. However, the impacts of global climate change, along with biotic and abiotic stressors, have led to significant reductions in wheat yields, posing a substantial threat to food security in the face of a rapidly growing population⁶. Wheat, which serves as a dietary staple for approximately 40% of the world’s population², is cultivated on more than 240 million hectares, with a global production volume of approximately 766 million tons - making it the second most-produced cereal worldwide⁷. Nonetheless, increasing climate variability, crop diseases, and demographic pressures continue to challenge the sustainability of production systems. One of the most significant threats to wheat production is foliar diseases, which affect up to 90% of global wheat cultivation areas (Chai et al., 2022). Among these, brown rust (Puccinia triticina), yellow rust (Puccinia striiformis), and tan spot (Pyrenophora tritici-repentis) are the most economically damaging diseases. Particularly in southern Russia, such diseases can lead to severe yield losses ranging from 50% to 70%^7,8. An estimated 62 million tons of annual wheat losses due to diseases equate to the nutritional needs of approximately 173 million people⁹. Conventional disease detection methods largely rely on manual visual inspection, which is time-consuming, subjective, and prone to errors^10,11. Moreover, symptoms similar to those caused by nutrient deficiencies or pest damage often result in misdiagnoses and unnecessary chemical treatments. Although laboratory tests provide more accurate results, they are typically costly and slow⁸. With the advancement of deep learning technologies, computer vision applications in agriculture have become increasingly widespread for disease and pest detection. While traditional computer vision methods rely on hand-crafted features such as color, shape, and texture, deep learning models are capable of learning feature representations automatically, resulting in more efficient and generalizable outcomes¹². In recent years, deep learning architectures such as Convolutional Neural Networks (CNNs) have demonstrated remarkable success in detecting wheat diseases, achieving accuracy rates exceeding 90%, often surpassing human performance^13,14. These models analyze infected leaf images and autonomously extract key features such as color patterns, textures, and structural deformations - enabling end-to-end learning without the need for manual feature engineering¹². However, the widespread adoption of these technologies faces several challenges. The lack of large-scale, balanced, and well-annotated datasets significantly limits model performance. Additionally, real-world conditions such as variable lighting, complex backgrounds, and low-resolution imagery adversely affect prediction accuracy^15,16. Early detection of wheat diseases not only helps prevent significant yield losses but also reduces unnecessary pesticide use, thereby minimizing environmental contamination and the risk of toxic residues in food. According to the Food and Agriculture Organization (FAO), plant diseases are responsible for an estimated $220 billion in annual global economic losses¹⁵. With the global population projected to reach 10 billion by 2050, agricultural production must increase by at least 50% to meet future food demand¹⁴. In this context, the widespread implementation of AI-driven agricultural technologies is crucial - particularly for ensuring the sustainability of wheat production in developing regions. Deep learning-based systems, which are increasingly replacing conventional diagnostic methods, offer significant advantages in enabling early and accurate detection of plant diseases, thereby contributing to sustainable agricultural practices. However, effective deployment in real-world scenarios requires the development of diverse, high-quality datasets and adaptation of models to field conditions, as well as efforts to enhance farmer training and awareness.

In this study, an innovative hybrid deep learning model is proposed for the automatic classification of wheat leaf diseases. The approach aims to achieve high classification accuracy and computational efficiency under both controlled laboratory conditions and variable field environments by combining the strengths of Convolutional Neural Networks (CNNs) and Vision Transformer (ViT) architectures. The broader goal of this research is to enhance agricultural productivity, reduce dependency on pesticides, promote the adoption of digital agriculture solutions, and contribute to sustainable farming practices for global food security. The proposed architecture, named HybridViT, integrates the local feature extraction capabilities of CNNs with the global contextual representation power of ViTs. By incorporating the Convolutional Block Attention Module (CBAM), the model is capable of effectively capturing localized disease symptoms such as lesion patterns, color alterations, and venation structures, as well as analyzing the broader spatial distribution of diseases on the leaf surface. These multi-scale features are selectively emphasized through attention mechanisms and adaptively weighted before being passed to the classification head. To enhance contrast and improve the visibility of early-stage disease symptoms, Contrast Limited Adaptive Histogram Equalization (CLAHE) was applied to the brightness component of the images during preprocessing. The model’s performance was comprehensively evaluated using three distinct wheat leaf disease datasets obtained from diverse geographical and environmental conditions. The experimental results indicate that the proposed hybrid model not only delivers high classification performance but also demonstrates strong generalization ability under varying real-world scenarios.

Motivation and novelty

Wheat is one of the cornerstones of global food security; however, in high-risk agricultural regions, leaf diseases can cause yield losses ranging from 50% to 70%, resulting in an estimated annual economic loss of approximately 220 billion USD. Traditional disease diagnosis methods, which primarily rely on manual inspection or laboratory testing, are inefficient in terms of time and labor, and operationally unsustainable for large-scale agricultural practices. Although deep learning-based approaches have demonstrated promising potential for automatic plant disease identification, their applicability under real-world field conditions continues to face several fundamental limitations. These challenges include limited generalization capacity in the presence of varying lighting conditions, structural noise, and complex backgrounds inherent to natural field environments; trade-offs between local and global feature extraction, where Convolutional Neural Networks (CNNs) excel at detecting fine-grained symptoms but struggle to capture long-range contextual dependencies, while Vision Transformers (ViTs) effectively model global relationships but often fail to detect subtle visual details; and a scarcity of high-quality annotated data, particularly for rare disease categories in agricultural imaging. In addition, existing methods often suffer from an imbalance between local and global feature extraction. CNN-based models are effective in capturing fine-grained symptoms but fail to model long-range dependencies, whereas Vision Transformers (ViTs) capture global context but are less sensitive to subtle visual cues. Limited availability of high-quality annotated datasets further restricts model generalization, particularly for rare disease classes.To address these structural limitations and to develop highly accurate and generalizable models, more holistic and balanced approaches are required. The HybridViT model introduced in this study is a novel hybrid architecture specifically designed to fill these gaps. The core contributions of the model are as follows:

1. Bidirectional feature extraction: ConvNeXt extracts fine-grained local disease patterns such as lesions and color variations, while ViT models global spatial dependencies across the leaf surface.

2. Adaptive feature integration: A gated fusion mechanism dynamically balances local and global representations, enabling input-dependent feature selection and improved feature complementarity.

3. Attention-based refinement: CBAM is applied after fusion to enhance discriminative feature learning through adaptive channel and spatial weighting.

4. CLAHE-based enhancement: CLAHE is applied to the brightness channel in preprocessing to improve low-contrast symptom visibility while preserving color consistency.

The model achieves 100% classification accuracy in controlled experimental datasets and reaches 99.10% accuracy on complex real-world field images, demonstrating its robustness even under varying environmental and visual noise conditions. By enabling early and precise disease detection, HybridViT supports reduced pesticide usage and promotes precision agriculture. In this regard, the proposed approach not only contributes to academic research but also holds strong potential to evolve into a scalable and deployable diagnostic solution for field-level agricultural applications.

Literature review

Accurate and early detection of diseases on wheat leaves is of paramount importance for enhancing agricultural productivity and preventing the spread of infections. In this context, image processing techniques combined with artificial intelligence have been extensively investigated to enable automated identification and classification of disease symptoms on leaf surfaces. Advances in deep learning models have facilitated precise and efficient recognition of such symptoms, thereby supporting timely intervention and mitigating yield losses.

In recent years, deep learning-based approaches for wheat disease detection have evolved significantly, moving from traditional CNN architectures toward more advanced hybrid, attention-based, and transformer-integrated models. In this evolutionary process, traditional machine learning and optimization algorithms exhibit several fundamental limitations. In particular, in cases where the search space is large or multimodal, early convergence, loss of diversity in subsequent generations, and sensitivity to parameter settings can negatively affect their performance¹⁷. In contrast, deep learning-based methods offer more powerful optimization strategies to overcome these issues. Particularly, recent studies (2021–2025) have demonstrated that lightweight CNNs, ensemble learning strategies, and attention-enhanced networks achieve improved robustness under real-field conditions, where illumination changes, background complexity, and image noise are common challenges. Moreover, Vision Transformer (ViT)-based and hybrid CNN–Transformer architectures have recently gained attention due to their ability to model long-range dependencies and global contextual relationships in plant images. However, existing studies indicate that pure transformer models often struggle with fine-grained feature extraction, while CNN-based methods may lack sufficient global context understanding, highlighting the need for more balanced hybrid frameworks.

Goyal et al. (2021)⁴ proposed a deep convolutional network architecture evaluated on the LWDCD2020 dataset, achieving high test accuracy. Their model outperformed widely used architectures VGG16 and ResNet50 by 7.01% and 15.92%, respectively, highlighting the potential of optimized deep learning architectures in disease classification. Kumar et al. (2023)¹⁸ introduced a hybrid approach integrating region-based RCNN models with a CE-SVM classifier for stripe rust classification. This method improved classification accuracy and model interpretability through precise disease region localization. Qushtom et al. (2025)¹⁹ developed a lightweight CNN model integrated with explainable artificial intelligence (XAI) techniques, specifically LIME and SHAP, to analyze wheat leaf images collected from the Kotli region (Azad Kashmir). Their approach enhanced transparency of model decisions, underscoring the significance of interpretability alongside accuracy in agricultural applications. Xu et al. (2023)¹² presented the RFE-CNN model, which combines parallel CNN architectures, residual channel attention blocks (RCAB), feedback blocks (FB), and elliptical metric learning (EML). Evaluated on diverse open datasets, the model demonstrated strong classification performance and generalizability. Dong et al. (2020)²⁰ proposed a differentiated reinforcement-based convolutional neural network (DACNN) with branched layers designed to sensitively capture subtle disease features during backpropagation. Their model, trained on a custom eight-class dataset augmented with data enhancement techniques, showed improved adaptability. Arinichev et al. (2022)⁸ introduced a lightweight neural network capable of simultaneously detecting multiple diseases such as yellow spot, yellow rust, and brown rust, addressing the multi-label classification challenge using macro and micro averaged metrics. Guo et al. (2020)²¹ employed hyperspectral imagery combining spectral and textural features within an SVM framework to accurately detect yellow rust, achieving superior performance compared to traditional methods. Lu et al. (2017)¹⁶ developed a weakly supervised deep learning system, DMIL-WDDS, capable of detecting diseases and localizing infected areas from image-level annotated wheat images collected under open-field conditions. This system has been further adapted for real-time field diagnosis via mobile applications. Reis and Turk (2024)²² combined CLAHE preprocessing with ensemble learning strategies to classify wheat leaf diseases using the IDLF model, which incorporates transfer learning and 13 distinct deep architectures. Their integration of RegNetY080 and random forest classifiers yielded notable improvements in classification accuracy. Pan et al. (2022)²³ developed the WR-EL method, fusing multiple CNN architectures optimized by SGDR-S, which significantly outperformed individual CNN models in rust disease classification. Genaev et al. (2020)²⁴ leveraged Densenet-based convolutional networks on smartphone-acquired images to accurately classify leaf and stem rust diseases. Jiang et al. (2022)²⁵ comprehensively evaluated seven CNN architectures across field and controlled environment datasets, highlighting the impact of transfer learning, model complexity, and field imaging conditions on wheat leaf disease detection robustness. Su et al. (2020)²⁶ employed a ResNet-101 FPN-based dual Mask-RCNN approach for automated detection of wheat spikes and Fusarium head blight (FHB) infected regions under field conditions, demonstrating utility in supporting efficient selection of breeding candidates. Maqsood et al. (2024)²⁷ addressed class imbalance via GAN-based data augmentation and utilized a ViT-GNN architecture trained with meta-learning, achieving superior performance in multi-disease detection tasks. Nigam et al. (2023)²⁸ proposed an EfficientNet-based lightweight deep learning model trained on natural field images, offering a highly effective diagnostic tool with potential for mobile deployment. Mi et al. (2020)²⁹ introduced a CBAM-enhanced C-DenseNet model for classifying yellow rust severity, which improved feature extraction capabilities and exhibited strong potential in real-world agricultural applications. Bao et al. (2021)³⁰ developed SimpleNet, a lightweight CNN that successfully identified wheat head diseases in complex backgrounds, optimized for mobile device compatibility. Liu et al. (2024)³¹ proposed a deep learning method based on orientation annotations to detect wheat diseases with variable aspect ratios and random orientations. Their approach, which included automated label assignment and localization refinement, achieved a mAP of 60.8% on the WDF2023 dataset and demonstrated robust performance in real-field conditions. Abetu et al. (2026)³² proposed a lightweight CNN with Adaptive Channel Reduction (ACR) for wheat disease classification on resource-constrained devices. Their approach, which included structured pruning and INT8 quantization, demonstrated robust performance under field perturbations while enabling efficient smartphone deployment. Li et al. (2026)³³ proposed MNCM, a lightweight MobileNetV3-large based model integrating channel prior convolutional attention (CPCA) and multi-head self-attention (MHA) for wheat disease identification. Their approach, which included layered learning rate optimization and partial layer freezing, outperformed OverLoCK-B, ConvNeXt, and MobileMamba while maintaining a compact architecture suitable for edge devices. Balasundaram et al. (2025)³⁴ proposed a deep learning pipeline integrating ViT-MUnet + + for segmentation and ARMV2-LSTM with ERF-SSA optimization for wheat stripe rust detection and grading across multiple severity levels. Chilakarao et al. (2025)³⁵ proposed a hybrid domain-adaptive framework combining MobileNetV2 and a Gradient Reversal Layer (GRL) to learn domain-invariant feature embeddings, which are subsequently classified using a K-Nearest Neighbors (KNN) model to improve generalization and interpretability in multi-class wheat leaf disease detection under domain shift conditions. Hassan et al. (2024)³⁶ proposed an end-to-end deep learning pipeline for wheat yellow rust detection in complex natural backgrounds, integrating YOLOv8-based leaf localization with an unsupervised rotation algorithm, followed by comparative evaluation of segmentation models where U-Net achieved the best performance, and transformer-based classifiers, particularly Swin Transformer, for final disease classification. StripeRust-Pocket (Liu et al., 2025)³⁷ proposed an offline mobile deep learning framework for wheat stripe rust severity assessment, integrating a lightweight two-stage StripeRustNet (MobileNetV2–DeepLabV3 + for leaf segmentation and ResNet50–DeepLabV3 + for lesion segmentation) and estimating disease severity via lesion-to-leaf pixel ratio, with deployment on an Android application for real-time field phenotyping and automated result management.

While these studies demonstrate strong performance under specific conditions, most approaches either focus on lightweight CNN-based designs for efficiency (Abetu et al., 2026; Li et al., 2026) or on hybrid and transformer-based models for improved representation (Balasundaram et al., 2025; Hassan et al., 2024). However, these methods generally rely on fixed architectural designs or static fusion strategies, lacking a dynamic feature integration mechanism that can adaptively balance local and global representations under varying input conditions. Collectively, these studies demonstrate significant progress in wheat disease classification using deep learning techniques. A structured comparison of these approaches is provided in Table 1, highlighting differences in methodologies, datasets, and performance metrics.

Table 1.

Summary of literature review in table format.

Author(s)	Year	Method/Model	Dataset	Accuracy	Key Findings
Goyal et al.4	2021	Novel Deep Convolutional Network (21 conv + 3 fc layers)	LWDCD2020	97.88%	Outperforms VGG16 and RESNET50 on 10 wheat disease classes.
Kumar et al.18	2023	RCNN variants + CE-SVM (Gaussian kernel)	Github, Kaggle, UCI (2352 + 12,849 augmented)	93.60%	High accuracy via zone extraction and CE-SVM; aids timely fungicide use.
Qushtom et al.19	2025	Lightweight CNN with XAI (LIME, SHAP)	450 localized images (Kotli, Azad Kashmir)	100%	Achieves perfect accuracy and interpretability with limited data.
Xu et al.12	2023	RFE-CNN (attention + feedback)	Multiple datasets (CGIAR, LWDCD2020, etc.)	98.83%	Excels on 5 diseases but limited on multi-disease images.
Dong et al.20	2020	DACNN (Differential Amplification CNN)	Custom 8-class wheat leaf dataset	95.18%	Improves sensitivity; outperforms LeNet-5, AlexNet, ZFNet, Inception V3.
Arinichev et al.8	2022	Lightweight CNN for multilabel classification	Multi-label wheat leaf dataset	90%	Detects multiple diseases with expert-level accuracy; mobile-friendly.
Guo et al.21	2020	SVM + spectral & textural features	Hyperspectral wheat leaf images	95.8%	Combining spectral and textural features improves yellow rust detection.
Lu et al.16	2017	DMIL-WDDS (VGG-FCN-VD16 / VGG-FCN-S)	WDD2017 (9230 field images)	97.95% (VD16)	Weakly supervised model for disease localization and classification.
Reis & Turk22	2024	IDLF (Ensemble + Transfer Learning, 13 DLA models)	Wheat Disease Detection Dataset (2400 images)	99.72%	Ensemble and CLAHE integration yielded highest accuracy.
Pan et al.23	2022	WR-EL (VGG, ResNet, DenseNet ensemble) + SGDR-S	Wheat Rust Dataset	92%	Ensemble outperforms single CNNs by up to 32%, improves F1 scores.
Genaev et al.24	2020	DenseNet-based CNN	CIMMYT (876 images)	F1: 0.90, AUC: 0.98	High performance despite varied image quality and devices.
Jiang et al.25	2022	Inception-v3 (transfer learning)	PlantVillage, FWDI	92.5%	Full fine-tuning best; environment and image quality impact results.
Su et al.26	2020	Dual Mask-RCNN (ResNet-101 FPN backbone)	55 wheat genetic lines field images	98.81%	Automatic segmentation of spikes and disease areas in field conditions.
Maqsood et al.27	2024	Transfer GAN + ViT + GNN + MAML ensemble	Public wheat disease dataset (6 classes)	99.20%	Data augmentation and model fusion achieve high multi-disease accuracy.
Nigam et al.28	2023	EfficientNet B4	WheatRust21	99.35%	Outperforms traditional CNNs; suitable for mobile deployment.
Mi et al.29	2020	C-DenseNet	WSRgrading (stripe rust, 6 severity levels)	97.99%	Attention improves fine-grained severity grading.
Nigam et al.38	2021	Deep Learning (CNN, Keras)	2000 wheat leaf images	97.30%	AI-based yellow rust detection improves with larger datasets.
Bao et al.30	2021	SimpleNet (Lightweight CNN + CBAM + feature fusion)	Wheat ear disease images from natural scenes	94.10%	Effective in complex backgrounds; mobile compatible.
Liu et al.31	2024	Rotated-RepPoints + Auto Label Assignment & Localization	WDF2023 (3875 images, 4 classes)	mAP: 60.8%, mRecall: 73.8%	Better detection of rotated and variable aspect ratio diseases.
Abetu et al.,32	2026	Lightweight CNN + ACR + Pruning + INT8	Ethiopian field dataset (5 classes)	95.67%	Adaptive Channel Reduction and INT8 quantization enable efficient smartphone deployment under field conditions.
Li et al.33	2026	MNCM (MobileNetV3-large + CPCA + MHA)	Self-built + public (7493 images, 5 classes)	97.49%	Integrates CPCA for local feature refinement and MHA for long-range dependencies, outperforming OverLoCK-B, ConvNeXt, and MobileMamba with a compact edge-deployable architecture.
Balasundaram et al.34	2025	ViT-MUnet++ + ARMV2-LSTM + ERF-SSA	Wheat leaf dataset (healthy + stripe rust)	91.61%	Integrates ViT-MUnet + + for precise disease segmentation and ARMV2-LSTM with ERF-SSA optimization for severity grading across multiple levels.
Chilakarao et al.35	2025	MobileNetV2 + GRL + KNN	11,204 images, 12 classes	97.66%	Best trade-off between accuracy, generalization, and interpretability.
Hassan et al.36	2024	YOLOv8 + Unsupervised Rotation + U-Net + Swin Transformer	Custom dataset (584 images + augmented + external test images)	95.8%	Fully automated pipeline for wheat rust detection in natural environments; U-Net achieved the best segmentation performance, and Swin Transformer provided the highest classification accuracy, with effective background removal improving results on small datasets.
Liu et al.37	2024	(MobileNetV2-DeepLabV3 + + ResNet50-DeepLabV3+)	5,013 wheat stripe rust images + field validation (100 images)	Leaf segmentation MIoU: 98.65% Lesion segmentation MIoU: 86.08%	Proposes mobile-based offline disease severity system; achieves high accuracy with efficient 2-stage segmentation and reduces annotation time from 20 min to 3 min per image

Despite significant advances in the literature, several limitations can still be identified. First, a large proportion of existing studies rely either on purely CNN-based or purely Transformer-based architectures, which limits their ability to simultaneously capture fine-grained local texture information and long-range global dependencies. Second, many recently proposed hybrid models focus more on architectural complexity than on achieving a balanced trade-off between efficiency and interpretability, thereby restricting their practical applicability in real-world agricultural scenarios. Although Transformer-based models introduce higher computational complexity compared to pure CNN architectures, they are capable of capturing long-range dependencies and global contextual relationships that are often critical in plant disease classification tasks. The proposed HybridViT framework is designed to address these limitations by introducing an adaptive gated fusion mechanism that dynamically integrates local and global feature representations, while maintaining a balance between accuracy and computational efficiency. Third, although recent studies employ advanced deep learning techniques, only a limited number of works systematically evaluate model generalization across multiple heterogeneous datasets representing diverse environmental conditions and domain shifts. Finally, computational efficiency aspects, such as GFLOPs and parameter optimization, are often insufficiently addressed in the literature, which limits the deployment of such models in resource-constrained edge devices. Collectively, these limitations highlight the need for a more balanced, efficient, and generalizable framework, which motivates the development of the proposed HybridViT architecture.

Materials and methods

The generalization capability of the proposed model was systematically evaluated using three different dataseta including the dataset developed by Long et al. (2023)³⁹ for training and assessing deep learning models for wheat disease classification, along with two additional datasets comprising diverse wheat leaf disease categories collected under varying environmental conditions. This comprehensive evaluation allowed for a rigorous assessment of the model’s performance across heterogeneous scenarios. This multi-dataset validation framework facilitated robust assessment of the model’s efficacy in both controlled experimental setups and complex, real-world field environments. In the preprocessing phase, input images were first transformed into the HSV color space to augment disease symptom visibility. Subsequently, Contrast Limited Adaptive Histogram Equalization (CLAHE) was selectively applied to the Value channel, effectively enhancing contrast while preserving the images’ natural color fidelity. This step notably improved the delineation of early-stage pathological features. The HybridViT architecture integrates the ConvNeXt backbone, adept at extracting fine-grained local features, with a Vision Transformer (ViT) module designed to capture long-range global contextual dependencies. The ViT component divides each image into 16 × 16 pixel patches, which are linearly projected into embedding vectors, enabling efficient modeling of spatial relationships. Concurrently, the ConvNeXt branch leverages advanced convolutional innovations such as expanded kernel sizes, normalized activations, and deep residual layers to refine detailed visual representations. Furthermore, a Convolutional Block Attention Module (CBAM) is incorporated to dynamically recalibrate feature responses via channel-wise and spatial attention mechanisms thus enhancing the discriminative power of the fused representations. Model optimization was conducted utilizing a label-smoothed CrossEntropyLoss combined with a OneCycleLR learning rate scheduler to promote rapid convergence and mitigate overfitting. Comprehensive performance evaluation employed metrics including accuracy, precision, recall, and F1-score, ensuring a rigorous and multidimensional analysis of classification capability across heterogeneous data conditions.

ConvNeXt: a modern CNN architecture

ConvNeXt is a contemporary deep learning architecture that reimagines the traditional Convolutional Neural Network (CNN) design using modern architectural principles⁴⁰. Developed to preserve the competitiveness of CNNs against attention-based models such as Vision Transformers (ViT)⁴¹, ConvNeXt combines scalability, simplicity, and high performance. By integrating contemporary components from Transformer architectures - such as Layer Normalization, GELU activation, and large convolutional kernels - it enhances the model’s capacity to capture both local and global contextual features effectively⁴². ConvNeXt leverages 7 × 7 depthwise convolutions to model broader spatial contexts, thereby minimizing information loss, while maintaining computational efficiency through the use of depthwise separable convolutions. Replacing Batch Normalization with Layer Normalization improves training stability, and employing GELU instead of ReLU provides smoother and more effective learning dynamics. A typical ConvNeXt block begins with a 96-channel input and applies a 7 × 7 depthwise convolution followed by Layer Normalization. This is succeeded by a 1 × 1 pointwise convolution that expands the channel dimension to 384, a GELU activation, and another 1 × 1 convolution that reduces the channel count back to 96, forming an inverted bottleneck structure. Finally, a residual connection is established by adding the input tensor to the block output^43,42. As illustrated in Fig. 1⁴⁴, this architecture facilitates gradient flow and mitigates the vanishing gradient problem, thereby supporting the training of deeper networks. ConvNeXt retains the residual structure of ResNet while incorporating architectural advancements inspired by ViT and Swin Transformer. The training strategy includes AdamW optimization, data augmentation techniques, and regularization methods such as DropPath, which collectively reduce overfitting and enhance accuracy. Its superior generalization ability is evidenced by outperforming Swin Transformer in benchmark tasks such as COCO object detection and ADE20K semantic segmentation⁴⁵. The ConvNeXt family includes several variants optimized for different computational budgets. ConvNeXt-Base features approximately 88 million parameters and delivers high accuracy, whereas ConvNeXt-Small (around 50 million parameters) and ConvNeXt-Tiny (around 30 million parameters) are tailored for systems with limited resources⁴⁶. In conclusion, ConvNeXt represents a robust solution for modern computer vision tasks by merging the simplicity of convolutional architectures with the representational power of Transformers. In this study, three variants of ConvNeXt were employed in binary classification tasks, establishing an effective bridge between CNN and Transformer-based paradigms⁴⁷.

Fig. 1.

Basic block design of ConvNeXt architecture.

Vision transformer (ViT)

Advancements in deep learning have led to significant progress in fields such as computer vision and natural language processing. In particular, convolutional neural networks (CNNs) have been used as a fundamental method for many years by effectively learning local features in image data. However, the limited receptive field of CNNs has proven inadequate in modeling long-range dependencies⁴⁸. To address this need, the self-attention mechanism specific to natural language processing has been adapted for visual data⁴⁹. Self-attention enables the creation of powerful representations by considering the contextual relationships of each element with all others in an input. Initially proposed for machine translation, this mechanism created a major breakthrough in natural language understanding tasks with the BERT model, due to its ability to learn bidirectional contextual information⁵⁰. Dosovitskiy et al.⁵¹ adapted the groundbreaking Transformer architecture from natural language processing to image processing and developed the Vision Transformer (ViT) architecture, which features a Multi-Head Self-Attention (MHSA) mechanism capable of capturing long-range dependencies in computer vision tasks. This model provides a strong alternative to traditional Convolutional Neural Networks (CNNs) by dividing images into patches and analyzing the relationships among these patches using the self-attention mechanism. By adapting the Transformer architecture to image processing, Vision Transformer (ViT) models images by dividing them into fixed-size patches and modeling the relationships among these patches via self-attention⁵². In this way, both local and global contextual information can be effectively learned⁵³. ViT has produced groundbreaking results in numerous tasks such as visual classification, segmentation, and recognition by overcoming the limitations of CNNs in modeling long-range dependencies. However, the ViT architecture also has some limitations due to its need for large datasets and high computational resources. Therefore, different structures have been proposed in the literature to enhance and make ViT more efficient. Among these are models such as DeiT (Data-efficient Image Transformer)⁵⁴, CaiT (Class-Attention in Image Transformers)⁵⁵, ConViT (Convolutional Vision Transformer)⁵⁶, and Swin Transformer⁵⁴, each aiming to improve ViT’s core structure from different perspectives. These models aim to offer advantages such as training with less data, faster convergence, better general performance, or lower computational cost. The Vision Transformer (ViT) model is successfully used in many different computer vision applications such as Image Classification, Image Segmentation, Object Detection, Medical Image Analysis, Face Recognition and Pose Estimation, Video Analysis and Action Recognition, and Classification of Agricultural and Plant Diseases thanks to its flexible structure and powerful global attention mechanism⁵³. As shown in Fig. 2, ViT mainly consists of the Patch Embedding, Transformer Encoder, and MLP Head modules.

Fig. 2.

General structure of the Vision Transformer (ViT) model.

Patch embedding layer

The Vision Transformer model divides the input image into non-overlapping and equally sized small segments, namely patches⁵⁷. For example, a color image of size 224 × 224 pixels is divided into 196 patches, each 16 × 16 pixels in size. Since these patches contain three color channels, each is a multidimensional matrix of size 16 × 16 × 3. In the model, each patch is first flattened into a one-dimensional vector and then embedded into the Transformer’s hidden dimension (e.g., 768) through a linear projection. Thus, each patch is placed into a high-dimensional space through a linear layer⁵⁸. Each patch is projected into a high-dimensional vector space suitable for the operation of the model, and then, as specified in Eq. (1), the embedding is performed via a linear layer⁵⁹. Additionally, positional encodings are added to the embedding vectors to provide the model with spatial information.

1

Here, represents the class token, and denotes the patch vector.  E is the embedding matrix used in the linear projection of the patches, and each patch is transformed into the relevant dimension by multiplying with this matrix. Furthermore, since the Transformer architecture inherently lacks the ability to process positional information, learnable positional encodings are added to the patch embeddings. This allows the model to consider the spatial locations of patches in the image.

Transformer encoder module

Transformer layers (Eqs. 2–3) consist of Multi-Head Self-Attention (MSA) and Feed-Forward Network (FFN) components⁶⁰. Each layer first applies Layer Normalization (LN); then contextual relationships are modeled through the attention mechanism, followed by non-linear transformations via the FFN. Residual connections applied to the input enhance training stability.

2

3

Here, “LN” refers to Layer Normalization, and represents the output of the layer.

Multi-head self-attention (MSA)

Multi-Head Self-Attention (MSA), one of the core components of the Transformer architecture, is used to model long-range dependencies among visual patches. In this mechanism, each input patch is transformed into query  Q, key  K, and value  V vectors. The attention scores are calculated as follows⁵⁰:

4

Here, , where  d denotes the embedding dimension and  N represents the number of patches. The scaling factor is used to prevent large dot product values from negatively impacting model learning. Since single-head attention provides limited contextual representation, MSA computes attention in parallel across multiple heads. Each head is defined as:

5

The outputs of all heads are concatenated to obtain the final output:

6

Here, , and are learnable parameter matrices. This structure allows the model to simultaneously learn different contextual relationships.

Feed-forward network (FFN)

The Feed-Forward Network (FFN) typically consists of two consecutive linear layers, and the GeLU (Gaussian Error Linear Unit) activation function placed between these layers facilitates the learning of non-linear representations. This structure is expressed by Eq. (7):

7

Layer Normalization and residual connections are present at the input and output of both sub-layers. This structure allows the model to be deepened and trained stably.

MLP head module

The class token (CLS token) at the output of the Transformer encoder is used for classification through a linear layer and softmax activation function⁵³. During pretraining, this module may include more complex multilayer perceptron (MLP) structures; however, during fine-tuning, a single linear layer is typically preferred. The model’s final class prediction is obtained through the following linear transformation over the class token at the beginning of the sequence:

8

Here, represents the model’s classification output, and and are the learnable weight and bias parameters of the classification layer. This structure allows the entire visual content to be effectively classified through a single representation vector. The mentioned architecture enables Vision Transformer to both learn local visual features and model global contextual relationships, thereby overcoming the limited receptive field problem of traditional convolutional neural networks (CNNs) and offering high-accuracy visual classification performance.

Convolutional block attention module (CBAM)

The Convolutional Block Attention Module (CBAM) is an effective attention structure developed by Woo et al. in 2018 to enhance the learning capacity of convolutional neural networks (CNNs) by sequentially applying channel and spatial attention mechanisms⁶¹. CBAM increases model accuracy by focusing on the most informative channels and spatial regions in feature maps. This module is particularly effective in computer vision tasks such as image classification, object detection, and segmentation, as it enables the network to better recognize and prioritize critical features, thereby improving overall performance⁶². As illustrated in Fig. 3, CBAM consists of two main components: the Channel Attention Module (CAM) and the Spatial Attention Module (SAM). First, the CAM determines the importance of each channel by applying global average pooling and max pooling operations. The resulting feature descriptors are fed into a shared multi-layer perceptron (MLP), which outputs the channel attention map. The input feature map is then reweighted by this attention map. Next, the SAM refines the output from CAM by applying average and max pooling along the channel axis to extract spatial context. These descriptors are concatenated and processed through a 1 × 1 convolutional layer, followed by a sigmoid activation, to produce the spatial attention map. This map is then applied to the feature representation to produce the final attention-enhanced output⁶³.

Fig. 3.

The structure of convolutional block attention module (CBAM).

These attention mechanisms are specifically designed to address CNNs’ shortcomings in suppressing irrelevant features and focusing on semantically meaningful regions⁶⁴. One of CBAM’s key contributions is its ability to dynamically refine feature representations within convolutional blocks. While the channel attention mechanism highlights feature-rich channels, the spatial attention mechanism emphasizes meaningful positional information, allowing the model to suppress background noise and focus on task-relevant areas^65,66. The operational flow of the attention mechanism proceeds sequentially, with channel and spatial attention applied to the input feature map  F. In the first stage, the CAM generates a channel attention map , which is element-wise multiplied with the input to produce a channel-refined feature map ⁶⁷:

9

In the second stage, the SAM processes to generate a spatial attention map , which is again applied via element-wise multiplication to obtain the final attention-refined representation :

10

The entire process can be summarized as:

11

In this formulation,  F represents the input feature map, is the channel attention map applied to  F, resulting in the channel-refined feature map . Then, denotes the spatial attention map derived from , which, when applied, produces the final attention-enhanced feature map . The operator indicates element-wise (Hadamard) multiplication used throughout this process. These sequential attention mechanisms effectively emphasize important features at both channel and spatial levels, enabling the model to produce stronger feature representations.

Channel attention module (CAM)

The Channel Attention Module (CAM) is a mechanism that adaptively reweights feature maps along the channel dimension in convolutional neural networks. This module addresses the “what” question—i.e., which channels in the input image carry more informative content⁶⁴. In this structure, where each channel is treated as a feature detector, the input feature map is compressed spatially while retaining the channel dimension.

As shown in Fig. 4, CAM applies global average pooling and global max pooling in parallel to the input feature map to generate two different statistical summaries. These summaries are passed through a shared multi-layer perceptron (MLP) with a ReLU activation. The MLP first reduces the dimensionality from  C to  C/r and then expands it back to  C. A sigmoid activation function is applied at the output to produce the final channel attention map ^68,69:

12

Fig. 4.

The structure of channel attention module (CAM).

Here, and are the weight matrices of the MLP, and denotes the sigmoid function. The same weights are applied to both feature descriptors, allowing for parameter sharing and increased model efficiency. Finally, the channel attention map is multiplied element-wise with the input features, emphasizing the most informative channels⁷⁰. The combination of average and max pooling enables the model to consider both general activation levels and local salient features, resulting in a comprehensive analysis of channel importance. The shared weight design enhances consistency, while the learned channel weights help the model to automatically focus on discriminative features.

Spatial attention module (SAM)

The Spatial Attention Module (SAM), the second core component of CBAM, complements the CAM by focusing on spatial relationships in the feature map. While the CAM answers the “what” question, SAM addresses the “where” by localizing informative regions⁶⁴. As illustrated in Fig. 5, SAM takes the output of CAM and applies average and max pooling operations along the channel axis. The resulting 2D maps ( and ) capture spatial context across all channels and serve as essential spatial descriptors. These are concatenated and passed through a 7 × 7 convolutional layer to encode spatial dependencies⁶⁸. A sigmoid activation function is applied to the output to obtain the spatial attention map , which is then multiplied with the input feature map to highlight salient spatial regions. The process is defined as follows:

13

Fig. 5.

The structure of spatial attention module (SAM).

Where denotes a convolution operation with a 7 × 7 kernel, and is the sigmoid activation function. The main goal of the SAM is to enhance the model’s sensitivity to local contextual information by emphasizing informative spatial locations and suppressing irrelevant background content. Experimental results confirm that applying spatial attention after channel attention results in more effective feature refinement and improved overall model performance.

Preprocessing phase

Image resizing strategy for transformer-based architectures

In this study, a Vision Transformer (ViT)-based architecture, specifically vit_base_patch16_224, is employed. Since transformer models operate on fixed-size input sequences derived from image patches, it is essential that all input images are standardized to a uniform resolution - in this case, 224 × 224 pixels. Given that the original images in the dataset exhibit varying aspect ratios, direct resizing to square dimensions can result in distortion and potential loss of semantic content. To address this, an aspect ratio-preserving resizing strategy is applied, wherein the shorter side of each image is resized to 224 pixels while maintaining the original proportions. This yields a larger image with consistent scaling. Subsequently, a center crop of 224 × 224 pixels is extracted from the resized image. This approach assumes that the primary subject - typically the leaf containing visible disease symptoms - is centrally located. As such, this method ensures the preservation of diagnostically relevant visual features while providing a consistent spatial layout across all samples. Overall, this preprocessing step not only mitigates geometric distortions but also enhances the model’s ability to learn discriminative patterns by feeding it spatially consistent and semantically rich inputs. As illustrated in Fig. 6, the preprocessing pipeline consists of CLAHE-based contrast enhancement applied to the brightness channel, followed by center cropping and resizing, and final normalization using ImageNet statistics. Overall, this preprocessing step ensures spatial consistency across all samples, reduces geometric artifacts, and enhances the model’s ability to learn discriminative patterns by providing semantically rich and structurally coherent inputs for both convolutional and transformer-based feature extraction.

Fig. 6.

Preprocessing pipeline including CLAHE-based contrast enhancement, center cropping, resizing, and normalization.

CLAHE application in HSV color space

In the preprocessing stage of this study, Contrast Limited Adaptive Histogram Equalization (CLAHE) was applied to enhance the contrast of leaf images and make disease symptoms more distinguishable. Initially, the images were converted into the HSV (Hue-Saturation-Value) color space, where brightness is separated as an individual channel. CLAHE was then selectively applied only to the V (Value) channel. This targeted enhancement ensures that contrast improvement affects only the brightness component, preserving the integrity of hue and saturation, and thus maintaining natural color representation. As a result, finer visual details are accentuated without introducing color distortion. CLAHE is a widely adopted image enhancement technique that operates by dividing the input image into small contextual regions (tiles) and applying histogram equalization independently within each region⁷¹. To prevent over-amplification of contrast, especially in homogeneous areas, the algorithm includes a clipping mechanism that limits the histogram redistribution⁷². This technique has shown particular effectiveness in improving feature visibility in low-contrast regions while mitigating the risk of noise amplification or over-enhancement. As illustrated in Fig. 7, applying CLAHE to leaf images significantly improves the visibility of disease-related patterns such as spots, vein structures, and color gradients. Additionally, natural variabilities such as uneven lighting and shadowing are reduced, enhancing the robustness of the model under different illumination conditions.

Fig. 7.

An example of applying CLAHE to an image of a diseased wheat leaf.

Following CLAHE, the images were converted back to the RGB color space and transformed into a tensor format compatible with the PyTorch framework. Finally, each channel was normalized to have a mean of 0.5 and a standard deviation of 0.5. This normalization step facilitates more stable and efficient training by ensuring balanced weight updates during the optimization process.

Proposed methodology

The overall architecture of the proposed HybridViT model is illustrated in Fig. 8, and consists of five main components: (1) preprocessing pipeline, (2) hybrid feature extractors, (3) adaptive gated fusion, (4) CBAM-based refinement, and (5) classification head.

Fig. 8.

Schematic diagram of the HybridViT framework, highlighting the integration of adaptive gated fusion and CBAM-based refinement modules for enhanced feature representation.

Hybrid feature extractors

To robustly capture both fine-grained pathological patterns and global contextual dependencies in wheat leaf imagery, the proposed framework introduces a dual-branch hybrid feature extraction module, as illustrated in Fig. 9. The design integrates a convolutional neural network (CNN)-based branch with a Vision Transformer (ViT)-based branch, thereby leveraging their complementary representational capabilities for enhanced feature discrimination and robustness, as supported by recent advances in hybrid vision modeling⁷³. The local feature extraction branch is constructed using a ConvNeXt-Tiny backbone, which modernizes classical convolutional architectures through improved design principles while retaining strong inductive biases for spatial feature learning. This branch is particularly effective in modeling localized disease manifestations, such as rust pustules, necrotic lesions, and subtle chromatic distortions on leaf surfaces. Following hierarchical feature extraction, the resulting convolutional feature maps are aggregated via global average pooling to obtain a compact representation. Subsequently, a linear projection layer maps this representation into a 512-dimensional latent embedding space, producing the local feature vector . In parallel, the global feature extraction branch employs a ViT-Small architecture with a patch size of . The input image is partitioned into non-overlapping patches, which are then processed through a multi-head self-attention mechanism. This formulation enables the model to effectively capture long-range dependencies and spatial interactions between disease-relevant regions across the entire leaf structure. The classification token ([CLS]), which encodes the global semantic representation of the input image, is extracted from the final transformer layer and subsequently projected into the same 512-dimensional embedding space, yielding the global feature vector . By embedding both local and global representations into a shared latent space, the proposed architecture ensures dimensional consistency and facilitates effective feature interaction in subsequent fusion stages. This dual-branch strategy enables simultaneous preservation of fine-grained pathological details and global structural context, thereby establishing a strong and well-balanced feature foundation for the downstream gated fusion and attention-based refinement modules.

Fig. 9.

Detailed schema of the Hybrid Feature Extractors. The architecture consists of a ConvNeXt-based local branch and a ViT-based global branch, both mapped to a 512-D projection space to facilitate feature integration.

Adaptive gated fusion

To effectively integrate complementary local and global feature representations, the proposed framework employs an Adaptive Gated Fusion mechanism, as illustrated in Fig. 10. This module is designed to avoid naive feature concatenation and instead learns to dynamically control the contribution of each modality based on the input-dependent context. Let denote the local features extracted from the ConvNeXt-Tiny branch and denote the global semantic features extracted from the ViT-Small branch. Instead of treating both representations equally, the model learns a gating vector that adaptively regulates feature importance:

14

Fig. 10.

Schematic representation of the Adaptive Gated Fusion mechanism, illustrating the dynamic integration of CNN and ViT features through a learnable gating network.

where represents feature concatenation, is a learnable projection matrix,  bis the bias term, and denotes the sigmoid activation function. The gating vector produces element-wise importance weights for feature selection. The final fused representation is computed as:

15

where denotes element-wise multiplication. This formulation ensures that each feature dimension is adaptively selected from either the CNN or ViT branch, rather than enforcing a uniform fusion strategy. From an architectural perspective (as implemented in Fig. 10), this mechanism enables the network to dynamically prioritize local texture-sensitive features (e.g., lesions, rust pustules) or global contextual cues (e.g., spatial disease distribution) depending on the input sample. This input-adaptive behavior significantly improves representational flexibility and reduces information redundancy compared to fixed fusion strategies such as concatenation or averaging. Overall, the adaptive gated fusion strategy plays a critical role in balancing fine-grained and holistic feature information, forming a robust intermediate representation that enhances downstream classification performance in wheat disease recognition.

CBAM-based feature refinement

After the feature fusion stage, the resulting representation is further refined using a lightweight attention-based mechanism inspired by the Convolutional Block Attention Module (CBAM), as illustrated in Fig. 11. Unlike standard CBAM implementations that operate on spatial feature maps, the proposed design applies channel-wise attention directly on the compact fused feature vector, making it computationally efficient and well-suited for deep latent representations. In this module, the fused feature vector produced by the adaptive gating mechanism is passed through a learnable transformation that estimates the importance of each feature channel. These learned importance scores are then used to re-scale the original feature vector in an adaptive manner. This process enables the model to emphasize more informative feature dimensions while suppressing less relevant or noisy responses. From an architectural perspective, this refinement step plays a critical role in improving feature discriminability. It allows the network to better focus on disease-relevant cues such as subtle texture variations, color inconsistencies, and lesion patterns that are often difficult to distinguish between similar wheat disease classes. By enhancing the most informative feature channels, the model becomes more robust to intra-class variation and inter-class similarity. Overall, the CBAM-based refinement stage improves the quality of the fused representation by introducing adaptive feature recalibration, thereby contributing to more stable and accurate classification performance in the proposed hybrid framework.

Fig. 11.

Architectural detail of the Enhanced CBAM Refinement module, depicting the sequential application of channel and spatial attention mechanisms on the 512-D fused feature representation.

Here,

 C

Feature dimensionality, fixed at 512 across fusion and refinement stages.

 H,W

Spatial dimensions of intermediate feature maps used in spatial attention.

 r

Reduction ratio in the CBAM shared MLP, set to 16 for computational efficiency.

Classification head and training strategy

The refined feature representation obtained from the CBAM-based attention module is finally passed to a lightweight classification head designed to ensure both high representational capacity and effective regularization. The classifier is composed of a fully connected layer that reduces the feature dimensionality from 512 to 256, followed by a GELU activation function to introduce non-linearity. To mitigate overfitting and improve generalization, a dropout layer with a rate of 0.3 is applied. Finally, the output layer maps the 256-dimensional representation to the number of target classes , producing the final prediction scores. The proposed model is trained using the AdamW optimizer with a learning rate of , which provides stable convergence for the hybrid architecture. Cross-entropy loss is used as the optimization objective to supervise multi-class classification. To enhance computational efficiency and reduce memory consumption, mixed-precision training is employed throughout the training process. Model performance is evaluated using classification accuracy on both validation and test sets, ensuring a reliable assessment of generalization capability. Overall, the proposed HybridViT architecture integrates complementary convolutional and transformer-based feature extraction with adaptive fusion and attention-based refinement mechanisms. This design results in a robust and well-generalized framework capable of effectively handling wheat disease classification under complex and real-world agricultural conditions. The overall training procedure is summarized in Algorithm 1.

Algorithm 1.

Training hybrid vision transformer for wheat leaf disease classification.

Network architecture

The proposed HybridViT architecture is a multi-branch deep learning framework that integrates convolutional neural networks and transformer-based models to capture both local and global feature representations. The overall architecture is summarized in Table 2, where each component is described in terms of its functional role and output dimension. The model employs a ConvNeXt-Tiny backbone for extracting fine-grained local texture features, while a ViT-Small Patch16 module is used to model long-range dependencies and global contextual relationships. To ensure feature compatibility, both branches are projected into a shared 512-dimensional embedding space prior to fusion. Feature integration is performed using a gated fusion mechanism, which adaptively learns the relative importance of CNN and ViT features based on input characteristics. This allows the model to dynamically balance local and global information. The fused representation is further refined using a lightweight CBAM (Convolutional Block Attention Module), which enhances discriminative feature learning by emphasizing informative channels and suppressing less relevant responses. The refined feature vector is then passed to a multi-layer perceptron classifier, which produces the final disease prediction.

Table 2.

Model complexity summary.

Model Component	Description	Output Dimension	Notes
ConvNeXt Tiny	CNN-based feature extractor	768 → 512	Local texture learning
ViT-Small Patch16	Transformer-based global feature extractor	384 → 512	Global context modeling
Gated Fusion	Adaptive feature fusion	512	Learnable weighting
CBAM (Light)	Channel-wise attention refinement	512	Feature recalibration
Classifier Head	MLP classifier	256 → N classes	Final prediction layer

In terms of computational efficiency, the proposed model contains 50.88 million parameters and requires 8.70 GFLOPs per inference, as summarized in Table 3. The input image resolution is fixed at 224 × 224, ensuring compatibility with both CNN and transformer backbones while maintaining stable computational requirements. Overall, the HybridViT framework achieves a balanced trade-off between representational richness and computational cost by combining the inductive bias of CNNs with the global reasoning capability of Vision Transformers. Moreover, the reported computational cost indicates that the model remains within a practical range for modern edge-assisted or cloud-based agricultural systems.

Table 3.

Computational complexity.

Metric	Value
Parameters	50.88 M
FLOPs	8.70 GFLOPs
Input Size	224 × 224
Batch Size	32
Backbone Type	CNN + ViT Hybrid

To better contextualize the computational cost of the proposed model, a comparison with commonly used baseline architectures is provided (Table 4). Compared to standalone ConvNeXt-Tiny and ViT-Small models, the proposed HybridViT introduces a moderate increase in computational cost due to the dual-branch design. However, this increase remains within a practical range for modern deep learning applications. Despite integrating both CNN and Transformer components, the use of lightweight backbones ensures that the overall complexity (50.88 M parameters and 8.70 GFLOPs) remains comparable to existing hybrid architectures, while achieving improved classification performance. In addition, the model achieves an average inference time of 13.62 ms per image on an NVIDIA Tesla T4 (16 GB), demonstrating its suitability for near real-time processing in practical agricultural applications. This latency corresponds to a throughput of approximately 73 images per second, further supporting its applicability in real-time systems. These findings indicate that the proposed model maintains a favorable balance between accuracy and computational efficiency.

Table 4.

Computational comparison.

Model	Params (M)	GFLOPs	Accuracy (%)
ConvNeXt-Tiny	~ 28 M	~ 4.5	98.99
ViT-Small	~ 22 M	~ 4.6	98.54
HybridViT (Ours)	50.88 M	8.70	99.10

Dataset description

In this study, three distinct datasets collected under different environmental conditions were utilized to comprehensively evaluate the generalization capability of the proposed model. These datasets were selected to represent both real-field and synthetically augmented scenarios, thereby enabling a more robust assessment of model performance under diverse conditions.

The first dataset, as illustrated in Fig. 12(a), referred to as the Ethiopian Field Wheat Disease Dataset (EFWDD), was obtained from a publicly available Kaggle repository⁷⁴. It was collected under real field conditions from a wheat farm located in the Holeta region of Ethiopia and is designed to support the detection of two major wheat leaf diseases: Stripe Rust and Septoria. The dataset contains a total of 1,266 images distributed across three classes, including 102 healthy leaves, 208 Stripe Rust-infected leaves, and 97 Septoria-infected leaves. Representative samples from each class are presented in Fig. 12(a), demonstrating the visual variability and symptom characteristics of the dataset.

Fig. 12.

Sample images from the wheat leaf disease datasets: (a) EFWDD (Holeta, Ethiopia) representing real field conditions; (b) BAWDD (Bangladesh) representing augmented and balanced samples.

The second dataset, shown in Fig. 12(b), referred to as the Bangladesh Augmented Wheat Disease Dataset (BAWDD), was collected from wheat fields in Bangladesh and contains 1,603 original high-resolution images (1920 × 1080 pixels)⁷⁵. It includes five disease categories: Black Point, Fusarium Foot Rot, Healthy Leaf, Leaf Blight, and Wheat Blast. To address class imbalance and enhance dataset diversity, the original dataset was augmented using a set of transformations including rotation, flipping, scaling, shifting, and center cropping, as well as color-based adjustments such as brightness, contrast, and saturation variations. This process resulted in a balanced dataset of 1,000 images per class, forming a total of 5,000 images, which was directly used in this study. Example images from each class are presented in Fig. 12(b), highlighting the intra-class variability and inter-class visual similarities present in the dataset.

The third dataset, shown in Fig. 13, employed as the primary dataset in this study, referred to as the UK-Ireland Field Wheat Disease Dataset (UIFWDD), was developed by Long et al. (2023)³⁹ as a comprehensive image collection designed for the training and evaluation of deep learning models aimed at classifying wheat leaf diseases. The dataset comprises images of both healthy and diseased wheat leaves, collected during the summer of 2019 across various field and glasshouse locations in the United Kingdom and Ireland. The images were captured using a range of devices, including iOS and Android smartphones as well as digital cameras, under diverse lighting and weather conditions and from multiple angles. This approach was intended to enhance the model’s generalizability under real-world conditions. For each disease category, data were collected from at least four distinct locations, and only leaves exhibiting clear symptoms of a single disease were included and labeled accordingly. As illustrated in Fig. 13, the dataset is structured into five distinct disease categories and comprises a total of 19,159 images.

Fig. 13.

Representative images from each class of the primary dataset, including healthy and diseased wheat leaves. The samples demonstrate significant variability in lighting conditions, viewpoints, backgrounds, and symptom appearance, highlighting the complex.

The datasets used in this study were selected to ensure a comprehensive evaluation of the proposed model under diverse real-world agricultural conditions, including data scarcity, class imbalance, and large-scale multi-environment variability, as summarized in Table 5. This design enables a more robust assessment of model generalization across different domain shifts.

Table 5.

Dataset summary and characteristics.

Dataset Name	Type	Acquisition Condition	Description	Experimental Purpose
EFWDD	Real-field dataset	Collected under natural field conditions in Holeta region, Ethiopia with limited sample availability and inherent class imbalance	A small-scale real-world dataset containing wheat leaf images affected by Stripe Rust and Septoria diseases alongside healthy samples	Used to evaluate model performance under data scarcity and real-field variability conditions
BAWDD	Augmented dataset	Derived from high-resolution field images collected in Bangladesh and enhanced using geometric and color-based augmentation techniques	A balanced dataset consisting of five wheat disease categories after extensive augmentation (rotation, flipping, scaling, brightness/contrast adjustments)	Used to assess model robustness under class imbalance correction and synthetic data augmentation scenarios
UIFWDD	Large-scale multi-environment dataset	Collected across multiple field and glasshouse locations in the United Kingdom and Ireland under diverse environmental and imaging conditions	A large-scale dataset with 19,159 images covering five wheat disease categories captured using multiple devices and environmental conditions	Used as the primary benchmark dataset to evaluate model generalization capability under real-world multi-location variability

To further highlight the differences among the datasets, it is important to note that EFWDD represents a small-scale real-field dataset with significant class imbalance and limited sample diversity, making it challenging for deep learning models to generalize. In contrast, BAWDD provides a synthetically balanced dataset through extensive augmentation, ensuring class balance while relying on controlled transformation strategies. On the other hand, UIFWDD constitutes a large-scale, highly heterogeneous real-world dataset, characterized by substantial variability in illumination, background complexity, viewpoints, and acquisition devices.

Data augmentation

As illustrated in Fig. 14, the third dataset developed by Long et al. (2023)³⁹ and used in this study as the primary dataset (UIFWDD) exhibits a pronounced class imbalance in the distribution of wheat leaf disease images. In the figure, paired bars are presented for each class: the first bars represent the original number of images, while the second bars display the adjusted distribution obtained after the application of data augmentation and downsampling techniques. For example, in the original dataset, the Septoria class contains 7,053 images, whereas the Healthy class comprises only 2,274 images. Such an imbalance limits the model’s capacity to effectively learn from underrepresented classes, leading to class bias and consequently a decline in overall performance. To address this issue, data augmentation methods were employed to increase the number of samples in minority classes, while the number of samples in majority classes was reduced through downsampling, thereby achieving a more balanced class distribution.

Fig. 14.

Class distributions before and after balancing in the wheat leaf disease dataset (UIFWDD).

To balance the dataset, a variety of image augmentation techniques were employed, including horizontal and vertical flipping, random rotation, brightness and contrast adjustments, zooming and cropping, as well as color space transformations. These transformations not only increased the number of training samples but also introduced greater variability, thereby enhancing the model’s generalization capability. Following the augmentation process, each class was balanced to contain exactly 3831 images. A comparison of the image counts before and after augmentation is provided in Table 6. This balancing process ensured that the model treated each class equally during training and contributed to a more robust and unbiased learning outcome.

Table 6.

Class distribution: before and after data augmentation (UIFWDD).

Class Name	Original Number of Images	After Data Augmentation
BrownRust	2,502	3,831
Healthy	2,274	3,831
Mildew	2,942	3,831
Septoria	7,053	3,831
YellowRust	4,388	3,831
Total	19,159	19,155

Data splitting strategy

In this study, the dataset was randomly and stratifiedly split into training, validation, and test sets using a ratio of 70%, 15%, and 15%, respectively. A fixed random seed was used to ensure reproducibility of the experiments. The training set was utilized for model optimization, while the validation set was employed for hyperparameter tuning and overfitting prevention. The test set was kept completely unseen during training and used exclusively for final performance evaluation to ensure an unbiased assessment of the model.

Experimental results

Experimental setup

All experiments in this study were conducted in both cloud-based and high-performance computing (HPC) environments. Model training initially took place on the Google Colaboratory platform, followed by the use of the TÜBİTAK ULAKBİM TRUBA infrastructure to handle larger data volumes and more computationally demanding tasks. While Google Colaboratory provided a convenient environment for deep learning prototyping and rapid experimentation, TRUBA offered a significant performance advantage, particularly for tasks requiring high memory capacity and multi-GPU support. The hardware specifications of the platforms used are summarized in Table 7.

Table 7.

Computational platforms and hardware specifications.

Platform	GPU	CPU	RAM	Disk	Notes
Google Colaboratory	NVIDIA Tesla T4 (16 GB)	-	12.67 GB	~ 112 GB Disk Space	Freely accessible cloud platform
TRUBA	2 × NVIDIA Tesla P100 (16 GB HBM2 each)	2 × Intel Xeon Scalable 6148 @ 2.40 GHz (40 cores total)	384 GB	High-performance shared cluster	Node used: barbun-cuda for compute-intensive tasks

The model was developed using the Python programming language and the PyTorch deep learning framework. The training process was carried out using a custom model named HybridViT, which integrates ConvNeXt and Vision Transformer (ViT) architectures, and incorporates a Convolutional Block Attention Module (CBAM) to enhance representational capability. During training, the AdamW optimization algorithm was employed, with an initial learning rate of 0.00005, a batch size of 32, and 224 × 224 pixel input image resolution. The loss function used was cross-entropy with label smoothing, and the OneCycleLR learning rate scheduler was applied to dynamically adjust the learning rate throughout training. Model performance was continuously monitored during training, and the model weights were saved whenever the best validation accuracy was achieved.

Performance evaluation metrices

The performance of machine learning models is assessed in terms of accuracy and generalizability to identify both strengths and limitations. In this study, the proposed model was tested on previously unseen data, and its effectiveness was quantitatively evaluated using a set of standard classification metrics. One of the most commonly used evaluation tools is the confusion matrix, which quantitatively represents the model’s prediction capability based on four fundamental components: true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). Derived from these values, accuracy, precision, recall, and F1-score were calculated to assess the model’s performance at both overall and class-specific levels. Accuracy refers to the proportion of correct predictions among all predictions, while precision indicates the ratio of true positives among all instances predicted as positive. Recall measures the model’s ability to correctly identify actual positive instances. The F1-score, as the harmonic mean of precision and recall, provides a balanced evaluation metric, particularly effective in the context of imbalanced datasets. The formulas for each of these evaluation metrics are presented in Eqs. (16)–(19):

16

17

18

19

where TP denotes True Positives, TN True Negatives, FP False Positives, and FN False Negatives.

These metrics were used to comprehensively evaluate the reliability and robustness of the model, offering detailed insights into its performance - especially in distinguishing diseased wheat leaves from healthy ones.

Obtained results

The proposed HybridViT architecture, which integrates ConvNeXt and Vision Transformer (ViT) backbones with the CBAM attention mechanism, demonstrates high accuracy, strong generalization ability, and resilience to environmental variability in the classification of wheat leaf diseases. The model’s performance was thoroughly evaluated on three distinct datasets collected under diverse geographical and environmental conditions. As presented in Table 8, the Precision, Recall, F1-Score, and Accuracy metrics indicate consistent and robust results across all datasets.

Table 8.

Classification performance metrics (Precision, Recall, F1-Score, Accuracy) of the Proposed HybridViT model across three distinct wheat leaf disease datasets.

Dataset	Class	Precision	Recall	F1-Score	Support
EFWDD 74	Healthy	1.00	1.00	1.00	21
	Septoria	1.00	1.00	1.00	21
	Stripe_rust	1.00	1.00	1.00	21
	Accuracy			1.00	63
BAWDD 75	BlackPoint	1.00	1.00	1.00	150
	FusariumFootRot	1.00	1.00	1.00	150
	HealthyLeaf	1.00	1.00	1.00	150
	LeafBlight	1.00	1.00	1.00	150
	WheatBlast	1.00	1.00	1.00	150
	Accuracy			1.00	750
UIFWDD 39	BrownRust	1.00	1.00	1.00	576
	Healthy	0.99	0.99	0.99	576
	Mildew	0.99	0.99	0.99	576
	Septoria	0.99	0.99	0.99	576
	YellowRust	0.99	0.99	0.99	576
	Accuracy			0.99	2880

In the first dataset (EFWDD), which consists of 83 balanced samples across three classes (Healthy, Septoria, Stripe Rust), the model achieved 100% accuracy, precision, recall, and F1-score. This outstanding performance can be attributed to the ViT component’s capacity for modeling global contextual relationships, the ConvNeXt module’s effectiveness in detecting fine-grained disease symptoms, and the enhanced image contrast provided by CLAHE-based preprocessing. The ability to deliver such high performance despite the limited sample size reflects the model’s potential to generalize effectively in low-data scenarios.

In the second dataset (BAWDD), which contains 1,000 synthetically balanced images across five classes (Black Point, Fusarium Foot Rot, Healthy Leaf, Leaf Blight, Wheat Blast), the model again achieved perfect scores across all evaluation metrics. This success demonstrates the collaborative strength of the ViT and ConvNeXt modules in extracting robust visual features and the CBAM module’s role in emphasizing the most informative attributes. The model’s ability to accurately differentiate between visually similar disease categories highlights its strong capacity for learning complex, multidimensional patterns.

The third dataset (UIFWDD) represents a more complex and challenging scenario, consisting of 3,835 field-collected images classified into five categories (BrownRust, Healthy, Mildew, Septoria, YellowRust). On this dataset, the model achieved an overall accuracy of 99%, with F1-scores ranging between 0.99 and 1.00 for all classes. Minor variations observed in the Healthy and YellowRust categories may be attributed to visual similarities in natural conditions, but they do not indicate any significant decline in classification performance. These results confirm the model’s robustness against environmental noise, lighting variations, and background complexity.

Overall, the HybridViT architecture effectively extracts both local and global features from CLAHE-enhanced leaf images, substantially improving classification accuracy, precision, and generalization. The training process, optimized with label-smoothed CrossEntropyLoss and the OneCycleLR learning rate scheduler, enables faster convergence while reducing the risk of overfitting. These findings suggest that the proposed model is well-suited for both academic research and practical deployment in real-world agricultural settings requiring reliable disease diagnosis.

The performance of the proposed model was comprehensively evaluated using confusion matrices derived from three distinct wheat disease datasets collected under varying environmental conditions. These matrices provide direct insight into the model’s class-wise discrimination ability and error distribution, making them essential for an in-depth performance assessment.

The first dataset (EFWDD) is relatively small-scale and balanced, containing only three classes (Healthy, Septoria, and Stripe Rust). As shown in Fig. 15, the confusion matrix indicates perfect classification, with all samples correctly assigned to their respective categories. The absence of off-diagonal values demonstrates that the model can clearly differentiate between semantically distinct disease patterns. This result reflects the combined effectiveness of ViT’s capability to capture long-range contextual dependencies and ConvNeXt’s precision in extracting localized pathological features. Additionally, the use of CLAHE-based preprocessing has enhanced low-contrast regions, making subtle disease indicators more distinguishable.

Fig. 15.

Confusion matrix for EFWDD, perfect classification across all classes.

The second dataset (BAWDD) includes 1,000 synthetically balanced samples across five classes (Black Point, Fusarium Foot Rot, Healthy Leaf, Leaf Blight, and Wheat Blast). As illustrated in Fig. 16, the confusion matrix again shows zero misclassification across all categories, confirming that the model achieved 100% accuracy. Despite visual similarities among certain classes, the model successfully distinguishes multidimensional symptom patterns. This outcome highlights the critical role of the CBAM attention module in emphasizing the most relevant channel-wise and spatial features. The model’s flawless separation of visually similar categories further supports its robust decision-making capacity.

Fig. 16.

Confusion matrix for BAWDD, no errors observed between classes.

The third dataset (UIFWDD) is more complex and challenging, comprising 3,835 test samples collected under real field conditions and encompassing five disease categories (Brown Rust, Healthy, Mildew, Septoria, and Yellow Rust). As presented in Fig. 17, the model achieved perfect classification for the Brown Rust and Healthy classes, while minor misclassifications occurred in Mildew, Septoria, and Yellow Rust. For instance, a small proportion of Mildew samples (1%) were confused with Healthy, and a reciprocal 1% misclassification was observed between Septoria and Yellow Rust. These discrepancies can be attributed to structural similarities in symptoms and natural variations such as lighting and shadow effects in field images. However, the low magnitude of these errors indicates no significant degradation in class separation performance.

Fig. 17.

Confusion matrix for UIFWDD, high accuracy with minimal class confusion.

Overall, the analysis based on confusion matrices demonstrates that the HybridViT architecture is capable of drawing clear inter-class boundaries and exhibits strong adaptability to various environmental conditions.

Ablation study analysis

As shown in Table 9, the proposed full HybridViT configuration achieves the highest performance with a test accuracy of 99.10%, confirming the effectiveness of the integrated hybrid design. The ablation results demonstrate that each architectural component contributes to performance improvement, either significantly or incrementally. Removing the CBAM module results in a slight decrease in accuracy (from 99.10% to 98.99%), indicating that attention-based refinement provides consistent, albeit modest, improvements in feature discriminability. Notably, the ConvNeXt-only baseline achieves a similar accuracy (98.99%), suggesting that CBAM mainly enhances the hybrid feature representation, rather than providing standalone gains. The comparison between single-backbone configurations further highlights the importance of hybrid design. While the ConvNeXt-only model achieves strong performance (98.99%), the ViT-only configuration shows a larger performance drop (98.54%), indicating that local texture features play a more dominant role in this task, whereas global contextual modeling provides complementary benefits. Furthermore, the analysis of fusion strategies shows that the proposed adaptive gated fusion consistently outperforms static alternatives. Specifically, both average and concatenation-based fusion methods achieve 98.89% accuracy, which is lower than the proposed method by 0.21%, confirming the advantage of learnable and input-adaptive feature weighting. In addition, removing CLAHE preprocessing results in a performance decrease from 99.10% to 98.85%, indicating that contrast enhancement contributes to improved robustness, particularly under varying illumination conditions. Although the numerical improvements between configurations appear relatively small, such behavior is expected in high-accuracy classification scenarios. Importantly, the improvements are consistent and systematic across all configurations, demonstrating that the proposed architecture benefits from the synergistic integration of CNN-based local feature extraction, Transformer-based global modeling, adaptive fusion, and attention-based refinement, rather than relying on a single component.

Table 9.

Ablation study results of the proposed HybridViT framework.

Configuration	CNN	ViT	Fusion Strategy	CBAM	CLAHE	Test Accuracy
Proposed (HybridViT)	✔	✔	Gated Fusion	✔	✔	0.9910
HybridViT w/o CBAM	✔	✔	Gated Fusion	✘	✔	0.9899
ConvNeXt-Only Baseline	✔	✘	—	✘	✔	0.9899
ViT-Only Baseline	✘	✔	—	✘	✔	0.9854
Hybrid (Average Fusion)	✔	✔	Average	✔	✔	0.9889
Hybrid (Concatenation Fusion)	✔	✔	Concatenation	✔	✔	0.9889
HybridViT w/o CLAHE	✔	✔	Gated Fusion	✔	✘	0.9885

Validation study

To systematically evaluate the effectiveness and generalization capability of the proposed model, a comprehensive validation study was conducted. In addition to the primary dataset used in this work, an additional independent dataset⁷⁷ was incorporated to increase data diversity and enhance the model’s generalization ability. In the experimental design, a unified training set was constructed by combining two different datasets. In contrast, the validation and test sets were derived from entirely separate data sources to ensure a more realistic and challenging evaluation setting. Specifically, the validation and test data were composed of two merged datasets: one containing the BrownRust, Healthy, and YellowRust classes⁷⁸, and another containing the Mildew and Septoria classes⁷⁹. This configuration introduces a significant domain shift between training and testing distributions, enabling a rigorous assessment of the model’s robustness under cross-domain conditions. Under this challenging evaluation scenario, the proposed HybridViT model achieved a test accuracy of 96.07%, demonstrating strong generalization performance across heterogeneous data distributions. The confusion matrix analysis (Fig. 18) indicates that most classes are correctly classified, with misclassifications primarily occurring between visually similar disease categories. This behavior is expected in fine-grained plant disease classification tasks due to subtle inter-class morphological similarities. Overall, the classification report presented in Table 10 and the confusion matrix results in Fig. 18 jointly confirm that the proposed model maintains high performance even under significant distributional shifts. These findings highlight that the HybridViT architecture offers strong generalization capability and robust performance for real-world precision agriculture applications.

Fig. 18.

Normalized confusion matrix for HybridViT on the merged test set.

Table 10.

Classification performance report.

Class	Precision	Recall	F1-score	Support
BrownRust	0.89	1.00	0.94	580
Healthy	0.98	0.97	0.97	782
Mildew	0.95	0.99	0.97	554
Septoria	1.00	0.91	0.95	497
YellowRust	0.99	0.93	0.96	693
Accuracy			0.9607	3106
Macro Avg	0.96	0.96	0.96	3106
Weighted Avg	0.96	0.96	0.96	3106

Robustness analysis and cross-validation results

As shown in Table 11, the proposed HybridViT model was comprehensively evaluated using a 5-fold cross-validation strategy on the EFWDD⁷⁴, BAWDD⁷⁵, and UIFWDD (Primary Dataset)³⁹ datasets. On the EFWDD dataset, the model achieved near-perfect classification performance across all folds, resulting in a mean accuracy of 100% with zero standard deviation. Similarly, on the BAWDD dataset, the model obtained a mean accuracy of 99.92% ± 0.15, indicating highly stable performance across different data splits. In contrast, the evaluation on the UIFWDD dataset, which represents a more challenging and realistic scenario, provides a more reliable indication of the model’s generalization capability. The model achieved fold-wise accuracies of 0.9911, 0.9914, 0.9892, 0.9866, and 0.9914, resulting in an overall mean accuracy of 99.04% ± 0.22, demonstrating consistent performance with low variance across folds. The training process exhibits stable convergence behavior, where validation accuracy remains consistently within the range of 98.6% to 99.1%, indicating robust learning without significant performance fluctuations. The slight difference between test accuracy (99.10%) and cross-validation results (99.04% ± 0.22) is expected due to variations in data partitioning and further confirms the stability of the proposed model. It is important to note that the near-perfect performance observed on EFWDD and BAWDD is influenced by their dataset characteristics, including limited scale (EFWDD) and augmentation-based balancing (BAWDD). Therefore, these results should be interpreted as indicators of model stability under specific conditions, rather than definitive evidence of real-world generalization. In contrast, the UIFWDD dataset, which includes diverse real-field conditions, provides a more realistic evaluation scenario, where the model maintains strong and consistent performance. This confirms that the proposed hybrid architecture effectively generalizes across different data distributions and environmental conditions.

Table 11.

Comparison of the proposed model with different studies and models.

Ref., Year	Methods	Dataset Used	Number of Classes	Accuracy
Bao et al., 202130	CNN + CBAM + Feature Fusion	Wheat ear disease images (natural)	3	94.10%
Xu et al., 202312	RFE-CNN	Multiple (CGIAR, LWDCD2020)	5	98.83%
Reis and Turk, 202422	Ensemble (DenseNet201, InceptionV3, RegNetY080, Xception + CLAHE + EL)	Wheat Disease Detection Dataset	3	99.72%
Goyal et al., 20214	21 Conv + 3 FC layers	LWDCD2020	10	97.88%
Long et al., 202339	CerealConv	Author-collected dataset (UIFWDD)	5	97.05%
Yousafzai et al., 202576	Multi-stage CNN-based Ensemble Learning (Bagging EL)	D1 (Wheat Disease Images Dataset UIFWDD)	5	86.78%
		D2 (Kaggle Wheat Leaf Dataset)	3	86.78%
		D3 (Kaggle Wheat Disease Dataset)	3	99.16%
Proposed Method	HybridViT (ConvNeXt + ViT + CBAM)	EFWDD74	3	100%
		BAWDD75	5	100%
		UIFWDD (Primary Dataset)39	5	99.10%

Table 11.

Cross-validation accuracy of the proposed HybridViT model on different datasets.

Dataset	Fold 1	Fold 2	Fold 3	Fold 4	Fold 5	Mean ± Std
EFWDD74	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000 ± 0.0000
BAWDD75	1.0000	0.9962	1.0000	1.0000	1.0000	0.9992 ± 0.0015
UIFWDD39	0.9911	0.9914	0.9892	0.9866	0.9914	0.9904 ± 0.0022

Comparison study

As shown in Table 11, various deep learning approaches for wheat disease classification reported in the literature have achieved high accuracy across different datasets and class configurations. These methods include CNN-based models, attention mechanisms, and multi-architecture ensembles, each demonstrating notable success under specific conditions. However, factors such as dataset size, class distribution, and environmental variability directly influence the performance of these models. The proposed HybridViT model, by effectively combining local and global feature extraction capabilities, has demonstrated superior accuracy and generalization across three distinct datasets with varying scales and complexities. While achieving perfect accuracy on smaller, balanced datasets, the model maintained consistent performance exceeding 99% even on more challenging datasets that reflect natural field conditions. This performance can be attributed to the proposed adaptive fusion of local and global representations, which enables more effective feature integration compared to conventional static fusion strategies. Additionally, the CLAHE-based preprocessing method improved the visual prominence of disease symptoms, thereby supporting the model’s ability to detect early-stage signs. Studies employing the same dataset as used in this work indicate that Long et al. (2023), who evaluated the CNN-based CerealConv model, reported an accuracy of 97.05%, whereas Yousafzai et al. (2025) achieved an accuracy of 86.78% using a multi-stage CNN-based ensemble learning approach (Bagging EL) on the dataset referred to as D1. In contrast, the proposed HybridViT model attains a superior accuracy of 99.10%. These findings highlight the effectiveness of jointly leveraging local and global feature representations.

Limitations and future directions

Despite the strong performance of the proposed HybridViT architecture, several limitations should be acknowledged. First, although three different datasets were used, two of them involve balanced and partially augmented data, which may contribute to higher classification performance under relatively controlled conditions. While the third dataset reflects real-world field variability, further validation across additional geographical regions, seasonal variations, and diverse acquisition devices is required to strengthen generalization claims. Second, the model was trained and evaluated using RGB imagery only. The incorporation of multispectral or hyperspectral data could provide richer feature representations and improve discrimination between visually similar diseases, such as Septoria and Yellow Rust. Third, although high accuracy was achieved, the inclusion of transformer-based components introduces additional computational overhead. This may limit direct deployment on resource-constrained edge devices without further optimization.

Future work will explore model compression, pruning, and lightweight variants to improve deployment efficiency. Finally, while the model demonstrates robustness to illumination variations through CLAHE preprocessing, extreme real-world conditions—such as severe occlusion, heavy shadowing, and mixed infections on a single leaf—were not extensively analyzed and remain open challenges. Future research will focus on expanding dataset diversity, incorporating multimodal sensing approaches, improving computational efficiency, and evaluating the model in real-time field deployment scenarios, to further enhance its practical applicability.

Conclusion

Wheat leaf diseases pose a significant threat to global food security, highlighting the need for accurate and reliable automated diagnostic systems. In this study, a hybrid classification framework, HybridViT, was developed by integrating the ConvNeXt architecture for local feature extraction with the Vision Transformer (ViT) for global contextual modeling. This integration is further enhanced by the Convolutional Block Attention Module (CBAM), enabling effective channel-wise and spatial attention refinement. Unlike conventional hybrid CNN–Transformer approaches that rely on static feature concatenation, the proposed framework introduces an adaptive gated fusion mechanism, which dynamically balances local and global feature representations. This design enables the model to effectively capture both fine-grained disease symptoms (e.g., lesions and color variations) and their spatial distribution patterns. The model was evaluated on three datasets collected under diverse conditions, including real-field and augmented scenarios. It achieved very high classification performance, reaching up to 99.10% accuracy on the most complex real-world dataset, while cross-validation results (99.04% ± 0.22) demonstrated stable and consistent performance across different data splits. It is important to note that near-perfect performance observed on smaller or augmented datasets is influenced by their structural characteristics, whereas results on large-scale real-world data provide a more reliable assessment of generalization capability. In addition, the model maintains a reasonable computational cost (50.88 M parameters and 8.70 GFLOPs) by combining lightweight CNN and Transformer backbones, making it suitable for modern edge-assisted or cloud-based agricultural applications. Despite these promising results, certain limitations remain. The framework has not yet been validated in real-time deployment scenarios, and its performance under extreme environmental conditions or unseen disease types requires further investigation. Future work will focus on improving computational efficiency through model compression and optimization techniques, as well as extending the approach to multi-crop and multi-disease classification tasks. Overall, the HybridViT framework provides a balanced solution in terms of accuracy, generalization, and computational efficiency, offering a promising direction for the development of practical AI-driven tools in precision agriculture.

Author contributions

N.A. and H.B. conceived and designed the study. N.A. developed the deep learning models, performed the experiments, and carried out data preprocessing and analysis. H.B. supervised the research, contributed to the methodological design, and provided critical revisions to improve the scientific quality of the manuscript. A.H.A.H. contributed to the interpretation of the results, validation of experimental findings, and critical review of the manuscript. E.A. assisted with dataset analysis, comparative evaluation, and manuscript editing. All authors discussed the results, contributed to the final manuscript, and approved the submitted version.

Funding

This work was supported by the Deanship of Scientific Research, Vice Presidency for Graduate Studies and Scientific Research, King Faisal University, Al-Ahsa, Saudi Arabia (Grant No. KFU262824).

Data availability

The datasets generated during and/or analyzed during the current study are available from the authors on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.