Explainable hybrid multi-branch CNN–ViT–GNN framework for robust hibiscus leaf disease classification Authors: Md Masum Billah, Saifuddin Sagor, Shahariar Hossain, Tahani Jaser Alahmadi, Mohammad Ali Moni, Mohammad Shorif Uddin Published: 2026 ABSTRACT AbstractEarly and reliable diagnosis of hibiscus leaf diseases is critical to protect horticultural yield. Yet, it remains challenging under real-time field conditions where uncontrolled lighting, clutter, and the non-contiguous nature of pathological symptoms blur diagnostic cues. To address these challenges, we introduce CNN-FusionViT-GNN. This explainable hybrid multi-branch framework synergizes the fine-grained texture extraction of a DenseNet201 backbone, the global contextual modeling of a Vision Transformer (ViT), and the relational reasoning of a Graph Neural Network (GNN). The model is trained and validated on ’Hibiscus,’ a curated field dataset of 1165 images from Bangladesh, which is strategically augmented to 8000 samples for robust training following a strict train-validation-test split. The proposed framework achieves a state-of-the-art accuracy of 98.33% with a macro F1-score of 0.98. The framework’s generalization is confirmed through high performance on external datasets: 98.78% accuracy on the 52-class Plant City dataset and 83.88% on the 10-class Tomato Leaf Disease dataset, while maintaining a rapid inference time of 10–45 ms. Furthermore, a multi-faceted Explainable AI (XAI) audit using LIME, Grad-CAM++, ViT Attention Maps, and Occlusion Sensitivity validates that the model’s decisions are driven by biologically meaningful symptom patterns rather than background artifacts. This study establishes a computationally efficient, transparent, and robust pathway for automated disease diagnosis in precision agriculture.Keywords: Agriculture, Horticulture, CNN, Feature extraction, Hibiscus leaf diseases, BioMedSubject terms: Computational biology and bioinformatics, Mathematics and computing Introduction Hibiscus plants, including Hibiscus rosasinensis and Hibiscus sabdariffa, are highly valued as both economically important and useful agricultural crops worldwide. Recent studies have reported their high medicinal value and potential for treating diabetes mellitus and steatotic liver disease, owing to their strong antioxidant activity1,2. In addition, roselle type is regarded as a delicate agricultural product both in the field of cultivation and in the post-harvest period3. Despite this importance, the agricultural industry faces serious threats from plant diseases, which destroy various parts of the hibiscus, reducing its growth and productivity4. As an example, new variants of tobamoviruses infecting hibiscus leaves and hibiscus bark have been identified, and they require specific diagnostic techniques5. The traditional method of manual inspection is subjective, labor-intensive, and time-consuming; therefore, it is not efficient for large-scale farming. This has led to the need to integrate automated Deep Learning (DL) methods to provide rapid and accurate disease classification and enable sustainable agricultural production6. The fact that CNNs have excelled at encoding hierarchical data and local textures in images has made them a pillar in this field7. CNNs have proved to be effective for visual classification tasks in natural scenarios8. However, CNNs perform better at extracting local features. Vision Transformers (ViTs) have recently been introduced and can be more efficient than traditional models, as they can capture long-range correlations and global contextual information through self-attention mechanisms9. Research in the agricultural sector has also shown that ViTs can significantly enhance crop productivity by enabling rapid diagnosis of leaf diseases10. The pixel-wise representation of images as patches also enables ViTs to capture the overall relationships among the components of an afflicted leaf, a capability that local filters in standard CNNs often fail to achieve. Furthermore, current automated systems often struggle in real-time environments where cluttered backgrounds and fluctuating illumination blur diagnostic cues. Our approach addresses these field-level challenges by integrating a robust preprocessing pipeline with a Multi-branch architecture that remains invariant to such environmental noise. Although recent advances in deep learning have introduced numerous CNN-based models for automated plant leaf disease detection, most existing approaches primarily rely on convolutional architectures that focus on local texture and color variations. While such local features are effective for identifying isolated lesions, they often fail to capture the broader spatial organization and relational structure of disease symptoms across the leaf surface11. In real-world Hibiscus pathology, disease manifestations are not merely independent spots; rather, they exhibit structured patterns in which the relative positioning, density, and co-occurrence of lesions convey critical diagnostic information12. Vision Transformers (ViTs) partially address this limitation by modeling global contextual dependencies through self-attention; however, they still treat image patches as sequential tokens without explicitly encoding relational interactions among semantically similar but spatially disjoint disease regions13. Consequently, neither standalone CNNs nor CNN–ViT hybrids are sufficient to model the complex topological relationships inherent in Hibiscus disease progression. To bridge this gap, a Graph Neural Network (GNN) becomes essential, as it enables explicit relational reasoning by modeling disease regions as interconnected nodes within a learned feature space. By capturing local texture characteristics, global contextual information, and inter-lesion structural relationships, the proposed Multi-branch framework, leveraging local textures (CNN), global context (ViT), and relational topology (GNN), addresses these critical limitations and provides a more holistic representation of leaf disease patterns14. The implementation of such multifaceted models in intelligent healthcare and agriculture requires Explainable Artificial Intelligence (XAI) to ensure that the decisions made by intelligent algorithms are reliable15. In the agricultural industry, XAI-based systems are considered essential to spur productivity and provide transparency16. Grad-CAM methods and attention visualization can help farmers and agronomists gain insight into how a prediction was made, relying on biologically meaningful patterns, thereby improving clinical reliability17. Moreover, XAI technology improves the transparency of crop recommender systems, allowing users to verify the soundness of the model logic based on geographical and operational parameters18. Despite technological advancements, numerous limitations remain, as the gap analysis of the available literature has identified. Existing technologies for identifying hibiscus leaf disease have limitations, including limited generalizability to other diseases and high computational cost19. Moreover, most existing models cannot efficiently handle noisy backgrounds or distinguish highly similar disease symptoms20. Although specific crops such as tea and tomato have been studied, there is a clear lack of specialized, high-performance hybrid models for the wide range of disease classes associated with the hibiscus plant21. This has created a pressing need for an architecture capable of achieving a good balance between representational power and computational efficiency to deploy edges in real time. To address these limitations, this study offers the following key contributions: Constructed a quality dataset called“Hibiscus”, which comprises a total of 1,165 high-quality JPEG images resized and standardized to 560 × 420 pixels in size. The images were all examined manually and annotated according to observed disease symptoms, yielding eight distinct classes: Citruspot, Early Mild Spotting, Fungal Infected, Healthy Hibiscus, Mild Edge Damage, Senescent, Slightly Diseased, and Wrinkled Leaf. Data were preprocessed using a stringent pipeline, including denoising, CLAHE, and gamma correction, and balanced through strategic data augmentation to ensure a statistically reliable training process and mitigate bias toward a particular class. Constructed a quality dataset called“Hibiscus”, which comprises a total of 1,165 high-quality JPEG images resized and standardized to 560 × 420 pixels in size. The images were all examined manually and annotated according to observed disease symptoms, yielding eight distinct classes: Citruspot, Early Mild Spotting, Fungal Infected, Healthy Hibiscus, Mild Edge Damage, Senescent, Slightly Diseased, and Wrinkled Leaf. Data were preprocessed using a stringent pipeline, including denoising, CLAHE, and gamma correction, and balanced through strategic data augmentation to ensure a statistically reliable training process and mitigate bias toward a particular class. We proposed a CNN-FusionViT-GNN framework that introduces a structured cross-modal interaction pipeline. It local texture-aware CNN embeddings guide the GNN’s relational reasoning via attention-based querying, while global semantic context from the Vision Transformer complements spatial and structural representations. This model multi-branch synergy enables the joint capture of fine-grained lesion textures, long-range spatial dependencies, and non-contiguous pathological patterns, which are critical for robust disease identification in complex field environments. We proposed a CNN-FusionViT-GNN framework that introduces a structured cross-modal interaction pipeline. It local texture-aware CNN embeddings guide the GNN’s relational reasoning via attention-based querying, while global semantic context from the Vision Transformer complements spatial and structural representations. This model multi-branch synergy enables the joint capture of fine-grained lesion textures, long-range spatial dependencies, and non-contiguous pathological patterns, which are critical for robust disease identification in complex field environments. Determined the statistical consistency and external validity of the proposed framework by extensive validation with stratified 5-fold and 10-fold cross-validation, which verified that the model is stable with small variance in a wide range of data partitions. External validation on secondary datasets, including a tomato 10-class dataset, further indicated that the DenseNet-FusionViT+GNN model performs strongly across a variety of imaging settings and plant species. Determined the statistical consistency and external validity of the proposed framework by extensive validation with stratified 5-fold and 10-fold cross-validation, which verified that the model is stable with small variance in a wide range of data partitions. External validation on secondary datasets, including a tomato 10-class dataset, further indicated that the DenseNet-FusionViT+GNN model performs strongly across a variety of imaging settings and plant species. Obtained a total Accuracy of 98.33%, which is higher than state-of-the-art CNN models, such as VGG19, Xception, InceptionV3, and DenseNet201. The model’s efficiency was also validated through GFLOPs analysis and inference latency measurements, ensuring real-time processing and enabling its use in any practical agricultural setup. Obtained a total Accuracy of 98.33%, which is higher than state-of-the-art CNN models, such as VGG19, Xception, InceptionV3, and DenseNet201. The model’s efficiency was also validated through GFLOPs analysis and inference latency measurements, ensuring real-time processing and enabling its use in any practical agricultural setup. Applied a multi-modal post-hoc interpretability based on LIME, Grad-CAM, and Occlusion Sensitivity to visually pinpoint the discriminative regions that affect model predictions. This transparency guarantees that the model is biologically relevant (based on the lesion patterns) and not image noise, and makes agronomists more confident and clinical disease diagnosis more reliable by automated diagnosis. Applied a multi-modal post-hoc interpretability based on LIME, Grad-CAM, and Occlusion Sensitivity to visually pinpoint the discriminative regions that affect model predictions. This transparency guarantees that the model is biologically relevant (based on the lesion patterns) and not image noise, and makes agronomists more confident and clinical disease diagnosis more reliable by automated diagnosis. The paper is organized as follows: “Related works” reviews related works, “Methodology” explains the methodology, “Result analysis” presents the results, “Discussion” provides the discussion, and “Conclusion” concludes the study. Related works Plant diseases continue to be a problem for the global agricultural sector, affecting crop yields, economic stability, and food security. Hibiscus plants, including those of Hibiscus rosa-sinensis and Hibiscus sabdariffa, are valuable plants because of their uses as decorative plants, and valuable functional crops with multiple medicinal value, as observed by Nogueira et al.22. Their antioxidant bioactive compounds are abundant, and their therapeutic use has been highlighted by recent pharmacological studies in the management of metabolic diseases such as diabetes and liver dysfunction (Li et al.23). Nonetheless, as Shafik et al.24 point out, several pathogens are often known to compromise the productivity of these crops. Arumuga et al.25 found that hibiscus plants are susceptible to diseases that cause damage to various parts of the plant, such as leaves and bark, which have serious effects in terms of reducing growth and yield. As a result, accurate diagnosis of these symptoms is essential to enable timely intervention and sustainable cultivation measures. The conventional methods for diagnosing plant diseases are largely based on manual observation by agricultural specialists, which, as Mutka et al.26 note, is inherently subjective, labor-intensive, and time-consuming. Experts are often too costly and time-consuming in a laboratory setting, leading to treatment delays. Although molecular techniques such as ELISA or RT-PCR offer high sensitivity, they are not scalable and thus cannot be used for real-time field monitoring Too et al.27. Computer vision has offered a revolution that gives an alternative to this, and automated systems can now be developed to rapidly and consistently classify diseases as explained by Tanzi et al.28. Convolutional Neural Networks (CNNs) have provided a solid foundation for image-based plant pathology. The works by Too et al.27 and Picon et al.8 confirm the effectiveness of CNNs in detecting hierarchical features and fine-grained local textures. Moreover, according to Billah et al.21, CNNs are highly effective for classification in the wild. The main weakness of conventional CNNs, however, is their strict localized receptive field, which in most cases cannot comprehend long-range spatial interactions and the global view of the overall leaf Hua et al.29. Vision Transformers (ViTs) have recently emerged in agricultural image analysis to address the drawbacks of localized feature extraction. In contrast to CNNs, ViTs use self-attention mechanisms to capture global contextual information, which is better than traditional models for tasks that use sequences (Liu et al.30). Balım et al.31 established that sophisticated ViT models are accurate at identifying diseases. In contrast, Li et al.23 proposed lightweight mobile ViT networks (PMVT) to detect them in real time on devices with limited resources. In addition to that, Wu et al.32 introduced edge-feature guidance modules to take advantage of ViTs to recognize insufficient local information, which is critical in early disease detection. Although ViTs are successful, representing the complex spatial and structural correlations among infected areas has proved challenging. Graph Neural Networks (GNNs) have become a potent method for capturing relational and structural dependencies by treating images as pixel node sets, as explored by Zhao et al.33 and Senior et al.34. According to REVATHI et al.35, GNNs can capture the relational structure among objects in an image and model natural dependencies. Thangamariappan et al.13 also demonstrated that hybrid architectures that integrate CNNs with GNNs can achieve much better performance by combining spatial locality with relational information. The most recent trend in the field is toward hybrid architectures that leverage the strengths of CNNs, ViTs, and GNNs, which are complementary. Shao et al.36 suggested that Transformers and CNNs can be used together to perform multi-focus image fusion. Li et al.19 also proposed an adaptive feature fusion model using CNN and GNN as parallel extractors to extract complex features. These hybrid frameworks address the weaknesses of single-method representation by offering a more complete feature representation Cao et al.17. Beyond architectural complexity, a significant limitation in existing plant pathology literature is the lack of model interpretability. Most frameworks operate as ’black boxes’, providing high accuracy without biological justification. Our work bridges this gap by applying a multi-lens XAI triangulation strategy, ensuring that the diagnosis is grounded in identifiable pathological features rather than artifacts. According to Fontes et al.37, XAI models play a critical role in ensuring algorithmic reliability. In the agricultural context, Martin et al.16. Naga Srinivasu et al.18 also claim that XAI-based frameworks, with methods such as Grad-CAM and LIME, can help experts identify biologically relevant patterns, thereby enhancing agronomists’ trust, as indicated by Pai et al.38. In39, a level-headed appraisal framework is presented that systematically contrasts the effectiveness of various network alignment methods. Synthetic and real-world biological experiments on biological datasets. In40, a graph attention and residual network-based GAT-HiC is proposed to predict three-dimensional chromosomal structure from Hi-C interactions. The authors41 develop a new Orbit-based GNN explainer (OExplainer) that combines both sample-level and method-level explainers on a pre-defined set of subgraphs. In such subgraph analysis, we aim to analyze graphs more comprehensively and in a more understandable way, and to assign an explainability score to each vertex for a given graph instance. Finally, while considerable progress has been made in deep learning applications for agriculture, there remains a significant research gap in developing specialized frameworks for Hibiscus plants. Available research often focuses on generic datasets or on various species, such as tea or tomato, as explored by Billah et al.7 and Solapure et al.42. Furthermore, Li et al.21, and Billah et al.23 indicate that there remain challenges in achieving high accuracy and lightweight networks for real-time mobile deployment. It is understandable that a Multi-branch fusion-based framework that combines CNN, FusionViT, and GNN is needed to address multiple Hibiscus disease symptoms with clinical-quality reliability. This method addresses existing constraints on transparency and provides a solid foundation for progress in Hibiscus plant pathology (Table 1). Table 1Summary of deep learning approaches for plant disease detection.Ref.ApproachDatasetAccuracyKey contributionLimitations4CNN with transfer learningHibiscus healthy/sick97.21%Focused on visual markers likeleaf discoloration and sporesRelies on local features;lacks global context21Custom residual networks (CRN)Image dataset51.0%Developed residual models for the Bangladesh tea industryAccuracy is insufficient for reliable field deployment24PDDNet lead voting ensemblePlantVillage 38 categories97.79%Integrated nine CNNs into a ensemble voting for species generalizationHigh computational cost due to multiple large networks27PMVT lightweight mobile ViTWheat, coffee, and rice93.6%Designed a mobile efficient ViT with 0.98M parameters and CBAMLower accuracy compared to larger hybrid architectures28Vision transformer (ViT)Femur fractures83.0%Evaluated ViT for clinical decision support in collaborative AIStruggles with small datasets and high feature similarity Summary of deep learning approaches for plant disease detection. Focused on visual markers like leaf discoloration and spores Methodology A comprehensive workflow for robust and explainable image detection has been proposed, a hybrid multi-branch model (CNN–FusionViT–GNN). To achieve consistent input distribution, data preprocessing begins with visual quality enhancement of raw images utilizing denoising, CLAHE, gamma correction, and normalization. To ensure a rigorous evaluation and prevent data leakage, the preprocessed dataset is first stratified into 70% training, 15% validation, and 15% testing subsets. Subsequently, strategic data augmentation—rotation, shifting, zooming, shearing, flipping, and brightness variation is applied exclusively to the training subset. This approach expands the training variety to reduce overfitting while ensuring that the model is validated and tested on original, unmodified real-world samples. The most effective CNN baseline model is DenseNet201, followed by InceptionV2, Xception, and VGG19. CNNs focus on local spatial patterns such as textures, colors, and edges, while accounting for long-range interdependence. Self-attention mechanisms help the Vision Transformer (ViT) model establish global contextual relationships. At the same time, the Graph Neural Network (GNN) represents extracted feature patches as interconnected graph nodes to capture higher-order structural dependencies. A hybrid fusion technique integrates these three modules using multi-head attention for cross-feature interaction and a fully connected SoftMax layer for classification. Standard metrics such as accuracy, precision, recall, F1-score, confusion matrix, and loss-accuracy analysis evaluate model performance. Multiple Explainable AI (XAI) approaches provide different insights into the model’s decision-making process to ensure interpretability. Model-agnostic, perturbation-based LIME (Local Interpretable Model-Agnostic Explanations) approximates the model locally with an interpretable surrogate to explain predictions. Grad-CAM and Grad-CAM++ highlight class-discriminative regions in CNN feature maps. ViT shows how the model focuses on picture regions based on Attention maps. Occlusion Sensitivity masks image regions to quantify prediction confidence and identify the most influential locations. Finally, Heatmaps provide an aggregated pixel-level visualization of feature importance. These techniques are categorized into model-agnostic (LIME), gradient-based (Grad-CAM, Grad-CAM++), attention-based (Attention Maps), and perturbation-based (Occlusion Sensitivity, Heatmaps) XAI methods. Figure 1 represents the Workflow for image classification of Hibiscus leaf images. Workflow for image classification using a hybrid CNN, ViT, and GNN model. Dataset description The Hibiscus Leaf Diseases Classification Dataset was compiled through systematic field data collection conducted across multiple agricultural and botanical regions in Bangladesh, including Dhaka, Khulna, Rajshahi, and other major hibiscus-growing areas. Images were captured directly from hibiscus plants in their natural outdoor environments to ensure authentic variations in illumination, leaf orientation, and background complexity that closely reflect real-world agricultural conditions. Both smartphone cameras and DSLR devices were used during image acquisition to balance accessibility and image quality. Photographs were taken under diverse lighting conditions and from multiple angles to capture natural heterogeneity in leaf color, texture, and disease expression. The dataset comprises a total of 1165 high-quality JPEG images, each resized and standardized to 560 × 420 pixels for consistency.The diagnostic definitions of the disease categories were grounded in established plant pathological markers, focusing on symptomatic criteria such as lesion morphology, chlorotic patterns, and mycelial colonization. To ensure the diagnostic integrity and biological authenticity of the dataset, all images were manually reviewed and annotated based on visible symptoms, with the final labels being cross-validated by agricultural specialists and plant pathology experts. This expert validation confirmed by Professor Dr. M. A. Rahim, Head of the Department of Agricultural Science at Daffodil International University (DIU), Bangladesh, and Dr. Md. Monirul Islam, Chairman of the Department of Plant Pathology and Seed Science at Sylhet Agricultural University, Bangladesh, that the eight well-defined classes, Citruspot, Early Mild Spotting, Fungal Infected, Healthy Hibiscus, Mild Edge Damage, Senescent, Slightly Diseased, and Wrinkled Leaf, accurately represent clinical manifestations common to Hibiscus rosasinensis. Each class is stored in a separate folder following a standardized directory structure to facilitate preprocessing and model training. Figure 2 shows sample images of Hibiscus leaves, and Table 2 provides the clinical definitions and label-wise distribution of the dataset. Representative sample images of different Hibiscus leaf disease classes. Table 2Label-wise distribution of the Hibiscus Leaf Disease dataset.Leaf classOriginal imageAdded imageTotal imageDescriptionCitruspot4735271000Small, circular necrotic lesions with distinct boundaries and consistent green coloration in surrounding tissueEarly mild spotting2267741000Initial stages of leaf degradation characterized by minor chlorotic spots along the leaf laminaFungal1508501000Advanced irregular necrotic patches with visible fungal mycelia or brownish-yellow spore-like clustersHealthy1098911000Asymptomatic leaves exhibiting full turgidity and uniform chlorophyll distributionMild edge damage839171000Marginal necrosis or mechanical degradation restricted to the leaf boundaries, often caused by early environmental stressorsSenescent569441000Progressive yellowing (chlorosis) and dehydration of the leaf surface due to natural aging or nutrient reabsorptionSlightly diseased409601000Leaves exhibiting multiple scattered micro-lesions and early signs of vitality lossWrinkled leaf289721000Severe surface deformation and leaf curling, often indicative of viral infection or high fungal load affecting the leaf structureTotal116568358000https://data.mendeley.com/datasets/5bzy89brkv/1. Label-wise distribution of the Hibiscus Leaf Disease dataset. https://data.mendeley.com/datasets/5bzy89brkv/1. s Image preprocessing and augmentation Image preprocessing The preprocessing pipeline was specifically engineered to transform raw field-captured images into a standardized format optimized for deep feature extraction. To address inherent sensor noise and environmental artifacts, a denoising step was first performed using Gaussian Blur43. This low-pass filter was preferred over Median or Bilateral filters because it provides a smoother noise reduction while preserving the natural gradients of lesion boundaries, which are critical for the GNN’s relational modeling. Following denoising, Contrast Limited Adaptive Histogram Equalization (CLAHE)44,45 was utilized to enhance local contrast, as visualized in the third column of Fig. 3. Unlike standard Global Histogram Equalization (GHE), which can over-saturate images and destroy subtle pathological details, CLAHE operates on localized tiles to make faint symptoms distinct. To ensure uniform input distribution across the network, Normalization was applied to scale pixel intensities to the [0, 1] range, facilitating stable convergence of the self-attention and message-passing layers. Finally, Gamma Correction (γ = 1.2) was implemented to non-linearly balance the exposure46. This sequence ensures that structural details under shadows are restored, making the features detectable by the hierarchical CNN–ViT–GNN architecture while maintaining the visual consistency illustrated across all leaf classes in Fig. 3. Leaf images of various conditions after preprocessing techniques. Data augmentation To ensure a meticulous evaluation and prevent data leakage, the dataset was first split into a training set using the original 1165 images. After splitting the dataset, strategic data augmentation was applied exclusively to the training subset. This approach ensures that the model is validated and tested only on original, unmodified real-world samples, while the expanded training pool provides the necessary diversity for the model to learn invariant features without memorizing the test distribution. To improve the model’s generalization across uncontrolled agricultural settings and prevent overfitting47, a targeted suite of augmentations, including Horizontal/Vertical Flipping and Rotation ()48, Zooming (0.2) and Shifting (0.1)49, and Brightness Adjustment50 was applied. The selection of these specific ranges was grounded in a preliminary sensitivity analysis conducted during the experimental design phase. We evaluated three intensity levels: “Low,” “Moderate” (the selected ranges), and “High.” It was observed that “High” intensity transformations (e.g., rotation or zoom > 0.4) led to semantic distortion, where critical diagnostic markers, such as the unique morphology of necrotic lesion boundaries and fine-grained venation patterns, were either cropped or unnaturally warped, resulting in a decrease in training stability. Consequently, the “Moderate” ranges were found to be the optimal intensity that maximizes dataset diversity from the initial 1165 images to a robust pool of 8000 samples while strictly preserving the biological integrity and diagnostic authenticity of the symptoms. These specific techniques, as visually represented in Fig. 4, were selected because they directly simulate the physical variability encountered during real-time field inspections, such as wind-induced leaf orientation, varying camera distances, and fluctuating natural light. Through the application of these strategic techniques, the initial dataset of 1165 original images was significantly expanded by adding 6835 augmented samples. This process resulted in a final balanced dataset of 8000 images, where each of the eight disease classes contains exactly 1000 samples. The detailed label-wise distribution and the counts of added images for each class are summarized in Table 2. Other common techniques, such as Grayscale Conversion and Cutout, were intentionally omitted. Grayscale conversion was avoided because color-coded diagnostic cues are vital for accurate identification. Furthermore, masking methods like Cutout were not used to prevent the accidental removal of unique lesion clusters indispensable for high-precision diagnosis. Similarly, artificial noise injection was excluded to avoid distorting the fine-grained pathological textures. By focusing strictly on geometric and photometric transformations, the augmentation process ensures scale and spatial invariance while maintaining the biological authenticity of the hibiscus disease symptoms. Samples of leaf images after strategic augmentation techniques. Other common techniques, such as Grayscale Conversion and Cutout, were intentionally omitted. Grayscale conversion was avoided because color-coded diagnostic cues are vital for accurate identification. Furthermore, masking methods such as Cutout were not used to prevent the inadvertent removal of unique lesion clusters indispensable for high-precision diagnosis. Similarly, artificial noise injection was excluded to avoid distorting the fine-grained pathological textures. By focusing solely on geometric and photometric transformations, the augmentation process ensures scale and spatial invariance while preserving the biological authenticity of hibiscus disease symptoms. CNN-FusionViT-GNN development The proposed hybrid CNN-FusionViT-GNN architecture is a robust and interpretable deep learning framework for hibiscus leaf disease classification. It processes color leaf images through preprocessing steps, including normalization, resizing, CLAHE, gamma correction, and denoising, to ensure consistent input quality. Feature extraction occurs via parallel pipelines. A DenseNet201 backbone captures hierarchical features (), while a multi-scale CNN with 1 × 1, 2 × 2, and 4 × 4 kernels and global average pooling produces 768-D embeddings. High-level representations are further modeled using a Vision Transformer (ViT) with four layers to capture long-range dependencies and a Graph Neural Network (GNN) with two layers and four attention heads to encode inter-feature relationships, both of which generate 768-D feature representations. A multi-head attention-based fusion module integrates these complementary features into a unified 768-D embedding, effectively balancing local texture information, global semantic context, and relational structural cues. The resulting fused representation is then classified into eight disease categories using a fully connected layer followed by a softmax activation. This multi-stream design enhances robustness, mitigates overfitting, provides rich and discriminative feature representations, and supports model interpretability through various explainable AI (XAI) techniques. Unlike conventional CNN-ViT or CNN-GNN hybrid architectures that rely on simple feature concatenation or sequential fusion, the proposed CNN-FusionViT-GNN framework introduces a structured cross-modal interaction pipeline. Specifically, local texture-aware CNN embeddings guide the GNN’s relational reasoning via attention-based querying, while global semantic context from the Vision Transformer complements spatial and structural representations. This multi-branch synergy enables the model to jointly capture fine-grained lesion textures, long-range spatial dependencies, and non-contiguous pathological patterns, which are critical for robust disease identification in complex field environments (Fig. 5). Overview of the proposed hybrid model for leaf disease recognition. Feature extraction via DenseNet201 Let the input image be denoted as . We use DenseNet20151 as the backbone network to extract spatial feature maps, leveraging transfer learning52 to initialize the model with weights pre-trained on the ImageNet dataset. This approach is preferred as the dense connectivity pattern of DenseNet201 alleviates the vanishing-gradient problem and encourages efficient feature reuse across layers. The initial feature representation is given by: Where is the input RGB image, denotes the DenseNet201 feature extraction function, is the resulting feature map, and h, w, c represent the height, width, and number of channels of the feature map, respectively. Multi-scale CNN feature embedding To capture global semantic information distinct from the spatial tokens, we employ a multi-scale pooling strategy. We extract features from different depths of the network: Where and are global feature vectors extracted at high, medium, and low spatial scales, GAP represents the Global Average Pooling53 operation, denotes a 1 × 1 convolutional layer used for channel reduction, and AvgPool signifies the average pooling operation. These vectors are concatenated and projected to form a compact global CNN embedding: Where is the compact global CNN embedding, is a learnable projection matrix, [; ] denotes the vector concatenation operation, ϕ is the Gaussian Error Linear Unit (GELU) activation function, and d represents the dimensionality of the final embedding space. The multi-scale pooling strategy enables the network to capture complementary visual cues across different semantic levels: high-resolution features emphasize fine-grained lesion textures, mid-scale features encode regional discoloration and spot distributions, and low-resolution features summarize global leaf deformation and shape irregularities. This design is particularly effective for hibiscus diseases, in which symptoms manifest simultaneously across multiple spatial scales. Transformer tokenization and positional encoding To facilitate global context modeling via the Vision Transformer (ViT), we reshape the spatial feature map into a sequence of flattened patches. Unlike the global embedding , this step preserves spatial details: Where is the sequence of flattened patches, denotes the tensor transformation operation, N = h × w represents the total number of spatial patches, and d refers to the dimensionality of each patch embedding. Unlike conventional Vision Transformers that tokenize raw image patches, the proposed framework tokenizes semantically enriched CNN feature maps, ensuring that each token represents a meaningful pathological region rather than low-level pixel information. Where is the initial input sequence for the Transformer, is a learnable classification token, and P denotes the learnable positional embeddings added to maintain spatial information. Transformer encoder (global context modeling) The token sequence is processed through L = 4 Transformer encoder layers, adhering to the foundational architecture of the Vision Transformer (ViT)54. Each layer consists of Multi-Head Self-Attention (MSA) and a Feed-Forward Network (FFN) with Layer Normalization (LN): Where is the token sequence at layer l, MSA represents the Multi-Head Self-Attention mechanism, LN is Layer Normalization applied to ensure training stability, and FFN denotes the Feed-Forward Network. A depth of four Transformer encoder layers was empirically selected to balance representational capacity and computational efficiency, as deeper configurations yielded marginal performance gains while significantly increasing inference cost. Graph neural network (relational reasoning) While the Vision Transformer captures global context via self-attention, it processes image patches as a sequence, potentially overlooking the structural connectivity between spatially distant but semantically similar disease symptoms. To address this limitation, we model the leaf image as a dynamic graph .The application of Graph Neural Networks (GNNs) in modeling complex biological structures has shown significant promise. Recent works, such as the GAT-HiC framework by40, leverage residual graph attention networks to reconstruct intricate 3D biological relationships, which supports our rationale for using relational reasoning to identify disjointed disease patterns. Additionally,39 demonstrated that hyperbolic deep graph learning can effectively align biological graphs in heterophilic environments. These advancements highlight the importance of capturing higher-order structural dependencies, a principle we incorporate into our multi-branch architecture to distinguish between genuine pathological features and environmental noise. Dynamic graph construction via k-NN The output token sequence from the Transformer encoder, denoted as , serves as the set of node features . Unlike static grid structures (e.g., standard CNNs), we dynamically construct edges to capture semantic similarities rather than merely spatial proximity. We employ the k-Nearest Neighbors (k-NN) algorithm in the latent feature space, following the dynamic graph construction principles introduced in55. For each node , we compute the pairwise Euclidean distance to all other nodes and establish directed edges to its k nearest neighbors: Where is the neighborhood set for node , represents neighboring nodes, dist(· ) denotes the pairwise Euclidean distance in the latent feature space, and k is the predefined number of nearest neighbors. In this study, the number of nearest neighbors was set to k=8. This value was determined through empirical validation to ensure the optimal trade-off between structural connectivity and feature purity. Experimental analysis revealed that lower values (k < 4) resulted in sparse graphs that failed to link non-contiguous lesion patterns, whereas higher values (k > 12) introduced noisy connections to irrelevant background regions. Therefore, k=8 was selected to maximize diagnostic accuracy while maintaining computational efficiency. The k-NN graph is dynamically reconstructed at each training iteration using the updated latent features, enabling the model to adapt to relational structures as representations evolve. Significance in plant pathology: the dynamic graph formulation is indispensable for addressing the spatial disjointedness inherent in hibiscus diseases. Unlike standard CNNs or ViTs that process regions using fixed grids or sequences, the proposed GNN explicitly models the topological relationships among scattered lesions. For instance, in ‘Early Mild Spotting’, fungal pustules appear as isolated clusters across the leaf surface. By establishing graph edges based on feature similarity rather than spatial proximity, the model effectively links these distant yet semantically related symptomatic regions. This relational reasoning enables the network to distinguish between structured pathological patterns and random environmental noise (e.g., dust or debris), thereby ensuring robust identification in uncontrolled field settings. Graph attention update To aggregate features from the constructed graph, we employ a Graph Attention Network (GAT)56. This mechanism learns to assign varying importance to different neighbors, enabling the model to focus on the most relevant pathological relationships. The attention coefficient , indicating the importance of neighbor j to node i, is computed as: Where is the attention coefficient, a is a learnable attention weight vector, is a shared linear transformation matrix, T denotes the transpose operation, ‖ represents concatenation, and is the non-linear activation function. Where is the updated node feature, σ is the Exponential Linear Unit (ELU) activation function, and represents the residual connection from the input node features. For multi-head attention, outputs from individual attention heads are concatenated and linearly projected to maintain dimensional consistency. The residual connection in the graph attention update stabilizes training by preserving the original patch-level semantics, ensuring robustness to potential variations in the dynamic graph topology across training iterations. Cross-modal attention fusion and classification After extracting features from separate branches, a naive concatenation would fail to prioritize the most relevant information. To address this, we introduce a Cross-Modal Attention (CMA) mechanism that dynamically weights the relational features from the GNN based on the global context provided by the CNN. This module acts as the decision-making core of our architecture. Attention mechanism formulation This module uses a multi-head attention mechanism57 to dynamically weight the relational features from the GNN according to the global context provided by the CNN. Let the global CNN embedding be denoted as and the spatially refined GNN features be . We utilize as the Query (Q) to seek relevant details from the GNN features, which serve as both Keys (K) and Values (V). This formulation allows the global texture context to select the most critical structural patches from the graph: Where Q, K, and V represent the Query, Key, and Value matrices respectively, are the refined node features from the GNN, and are learnable projection matrices for the cross-modal attention mechanism. The global CNN embedding is employed as the Query in the cross-modal attention mechanism because CNN features explicitly encode disease-specific texture patterns, such as lesion boundaries and color irregularities. Using these features to guide attention ensures that the model prioritizes structurally relevant graph nodes associated with pathological regions, rather than relying solely on abstract global semantics from the Transformer CLS token. The attention scores are computed to determine which graph nodes (image patches) contribute most to the final classification: Where is the pooled relational feature vector, denotes the normalization function that converts attention scores into probabilities, and V are the Query, transposed Key, and Value matrices respectively, and is the scaling factor used to stabilize gradients. Crucially, since the Query Q is a single vector (1 × d), the resulting output is a condensed feature vector that aggregates the most significant disease symptoms from the entire graph, effectively filtering out background noise. Functionally, this operation acts as a context-aware, weighted pooling mechanism, in which the global CNN context dynamically determines how much information to aggregate from each graph node. Feature aggregation and prediction To form a holistic representation, we fuse three distinct feature sets: the attention-pooled graph features (), the global semantic context from the Vision Transformer’s CLS token (), and the original multi-scale CNN embedding (). This preserves information from local texture, global context, and relational structure simultaneously: Where is the unified multi-branch representation, LayerNorm denotes normalization applied after feature concatenation, and represents the global semantic context from the Transformer classification token. Layer normalization and dropout regularization58 are applied after feature concatenation to mitigate potential overfitting caused by high-dimensional feature fusion, particularly under extensive data augmentation. Finally, this unified feature vector is passed through a fully connected (Dense) layer with dropout for regularization, followed by a Softmax activation function to predict the probability distribution over the C disease classes: Where is the predicted probability distribution over the C=8 disease classes, ϕ is the GELU activation function, and are weight matrices, and are the corresponding learnable bias vectors. Table 3 presents the layer-wise architectural configuration and parameter distribution of the proposed hybrid CNN-FusionViT-GNN framework. The pipeline begins with a DenseNet201 backbone that extracts high-dimensional local feature maps (), which are subsequently projected into a 768-dimensional embedding space compatible with Transformer processing. These embeddings are reshaped into token sequences, augmented with a learnable classification ([CLS]) token and positional embeddings, and processed through four Transformer encoder layers employing Multi-Head Self-Attention (MHSA) to capture global contextual dependencies. To overcome the inherent sequential modeling limitations of Transformers, the resulting tokens are further structured as a graph and refined using stacked Graph Neural Network (GNN) layers, enabling explicit modeling of non-contiguous relational dependencies among feature patches. In parallel, a multi-scale CNN branch extracts complementary spatial features at multiple resolutions to preserve fine-grained textural information. As summarized in Table 3, the final decision module integrates the global [CLS] context, GNN-enhanced relational features, and multi-scale CNN embeddings via an attention-based fusion mechanism. The resulting 2304-dimensional fused representation is projected through a fully connected layer for final classification. The overall parameter count (approximately 60.2 million) reflects a balanced trade-off between representational capacity and computational efficiency, supporting practical deployment in resource-constrained agricultural monitoring scenarios. Table 3Layer-wise architecture and parameter configuration of the proposed Hybrid CNN-FusionViT-GNN model.Layer typeOutput shapeParametersConnections / descriptionInput layer0RGB input imageDenseNet201 backbone18,321,984Pre-trained DenseNet feature extractorHigh-resolution projection1,475,3281× 1 convolution on DenseNet featuresSequence reshaping49 × 7680Flattened spatial feature tokensCLS token addition50 × 7680Learnable classification token appendedPositional embedding50 × 76838,400Learnable positional encodingsTransformer encoder (4×)50 × 76823,625,728Multi-head self-attention and MLP blocksGNN layers (2×)50 × 7689,468,448Graph attention and feed-forward propagationMulti-scale CNN features23044,423,984Aggregated low-, mid-, and high-level CNN featuresCNN embedding FC7681,770,240Fully connected projection of CNN featuresFusion multi-head attention1 × 7682,363,268CNN query attending over GNN representationsFeature Concatenation23040Fusion output, CLS token, and CNN embeddingsFusion FC + dropout7681,770,240Dimensionality reduction and regularizationOutput layer86,152Softmax classification layerTotal parameters60,209,992 (229.68 MB)Trainable parameters59,989,952 (228.81 MB)Non-trainable parameters220,056 (884.75 KB) Layer-wise architecture and parameter configuration of the proposed Hybrid CNN-FusionViT-GNN model. Experimental setup and training parameters The experiments were conducted on the Lightning AI platform using an NVIDIA L4 GPU (20 GB VRAM) and the PyTorch deep learning framework to facilitate parallelized model training. All input images were standardized to a resolution of . This resolution was selected to ensure architectural compatibility with the pre-trained ImageNet weights of the DenseNet201 and ViT backbones, while balancing the need to capture fine-grained pathological textures and the need to keep inference latency low. To ensure reproducibility of the findings, a fixed random seed of 42 was used for weight initialization and data shuffling, and the model required approximately 240 seconds per epoch on this hardware setup. The optimized hyperparameter configuration for the hybrid CNN-FusionViT-GNN model is summarized in Table 4. Training was performed using the Adam optimizer,chosen for its robust adaptive learning-rate capabilities, which are particularly effective for hybrid architectures in high-dimensional feature spaces. An initial learning rate of was selected after a systematic grid search, as it provided the most stable stochastic convergence and effectively mitigated gradient oscillations. Categorical cross-entropy was used as the loss function, as it is the standard objective function for multi-class classification with mutually exclusive categories. To provide statistically reliable performance estimates, the diagnostic metrics reported in this study are computed as the average of results from stratified 5-fold and 10-fold cross-validation procedures, ensuring consistency across multiple model runs and independence from specific data splits. Furthermore, to mitigate overfitting and improve generalization, we integrated a dynamic regularization pipeline that combines an EarlyStopping mechanism (patience = 5) and a ReduceLROnPlateau scheduler (factor = 0.2, patience = 3). Although the maximum training limit was set to 30 epochs, the model consistently converged within 12–15 epochs, at which point the EarlyStopping mechanism terminated training to preserve the peak validation performance. This rigorous configuration ensures high reproducibility and establishes a robust baseline for the diagnostic performance analysis. Table 4Training hyperparameters employed for the proposed CNN-FusionViT-GNN model.HyperparameterValueRandom seed42Number of epochs30Batch size32Input image sizeInitial learning rateOptimizerAdam optimizer ()Loss functionCategorical cross-entropyEarly stoppingPatience = 5, monitored metric = validation accuracy,restore best weights enabledLearning rate schedulerReduceLROnPlateau (factor = 0.2, patience = 3,cooldown = 10)CallbacksEarlyStopping, ReduceLROnPlateau, ModelCheckpoint Training hyperparameters employed for the proposed CNN-FusionViT-GNN model. Multi-perspective interpretability framework To mitigate the inherent black-box nature of the proposed hybrid CNN-ViT-GNN architecture and ensure deployment-grade diagnostic reliability, we adopt a multi-perspective interpretability triangulation strategy. Relying on a single explainability technique often leads to biased or incomplete interpretations; therefore, we systematically cross-validate the model’s decision-making logic using three complementary lenses: (i) gradient-based localization (Grad-CAM and Grad-CAM++) to audit the CNN backbone’s sensitivity to disease-specific textures, (ii) attention-based visualization to interpret the Vision Transformer’s global contextual reasoning, and (iii) perturbation-based analysis (LIME and Occlusion Sensitivity59) to assess the causal relevance of spatial regions under input masking. Beyond standard localization, the integration of structural motifs in explaining biological GNN decisions, as explored by41, emphasizes the need for transparent reasoning. Our multi-perspective XAI strategy aligns with this requirement to ensure model predictions are grounded in biologically relevant features. This triangulated framework ensures that discriminative regions consistently align with biologically meaningful pathological symptoms (e.g., necrotic lesions, chlorosis, and edge degradation) rather than background artifacts, thereby establishing explainability suitable for high-stakes agricultural disease diagnosis. Local fidelity analysis using LIME Local interpretable model-agnostic explanations (LIME)60 is employed to evaluate the local faithfulness of the proposed hybrid model by approximating its behavior in the vicinity of a specific input image. Given an input leaf image x, LIME constructs an interpretable surrogate model g by minimizing the following objective: Where ξ (x) is the explanation model, argmin denotes the value that minimizes the loss, g is an interpretable model from the set of potential explanations , f is the complex black-box model, π is the proximity kernel, and Ω (g) represents the model complexity constraint. In the context of hibiscus leaf disease classification, LIME consistently highlights super-pixels corresponding to lesion boundaries and discolored regions while suppressing background vegetation and illumination artifacts. This confirms that the model’s predictions are locally driven by disease-relevant morphology rather than spurious contextual cues. However, as LIME captures only local surrogate behavior around a single instance, it does not provide spatial continuity or global consistency, motivating the use of gradient-based localization methods. Gradient-based feature localization (Grad-CAM and Grad-CAM++) To interpret the spatial focus of the CNN backbone, we employ gradient-based class activation mapping techniques (Grad-CAM)61. Grad-CAM computes class-specific neuron importance weights by globally averaging the gradients of the target class score with respect to the final convolutional feature maps : Where is the importance weight for feature map k and class c, is the prediction score for class c, is the activation value at position (i, j) of feature map k, and Z is the total number of pixels in the feature map.The corresponding localization map is obtained as: Where is the resulting visual class activation map and ReLU ensures only positive contributions to the class of interest are visualized. However, hibiscus diseases such as Early Mild Spotting often manifest as multiple, spatially scattered lesions, where standard Grad-CAM may yield incomplete localization. To address this limitation, we incorporate Grad-CAM++62, which leverages higher-order gradient information to assign pixel-level importance weights: Where represents the pixel-level importance weights for Grad-CAM++, utilizing higher-order derivatives (second and third-order gradients) to improve localization accuracy for multiple scattered lesions. By enabling fine-grained and multi-instance localization, Grad-CAM++ enhances the model’s ability to capture subtle and early-stage pathological patterns that are critical for timely disease detection. Global dependency analysis via ViT attention maps While CNN-based explanations excel at capturing localized texture variations, they are inherently constrained by limited receptive fields and may overlook global structural dependencies. To analyze long-range contextual reasoning, we visualize self-attention maps derived from the Vision Transformer (ViT) encoder layers. Unlike gradient-based explanations that primarily emphasize local texture responses, attention maps provide a holistic view of how spatially distant regions jointly influence the model’s decision. The multi-head self-attention mechanism computes the attention matrix as: Where A is the attention matrix representing relationship weights between image patches, Q and K are the Query and Key projections from the Transformer encoder, and is the normalization factor. In hibiscus pathology, disease indicators are often expressed as global cues, such as venation distortion, margin degradation, and leaf-wide chlorosis, rather than as isolated lesions. For example, distinguishing Mild edge damage from Healthy leaves requires understanding the interaction between boundary geometry and internal texture patterns. The attention visualizations confirm that the proposed hybrid architecture jointly attends to localized lesions and global structural attributes, thereby validating the complementary integration of CNN-based local feature extraction and ViT-driven global context modeling. While these explainability techniques provide complementary insights, they remain post-hoc approximations and do not guarantee causal explanations, motivating future exploration of inherently interpretable architectures. Result analysis The comprehensive evaluation of the proposed CNN-FusionViT-GNN framework was conducted through a multi-tiered validation strategy. This assessment aims to quantify the model’s diagnostic efficacy, statistical stability, and real-world generalizability. To provide a high-fidelity performance audit, we analyzed the model using class-specific metrics, holistic visualization via confusion matrices and ROC curves, and rigorous cross-validation techniques. This section analyzes the empirical results to demonstrate how the synergistic integration of local, global, and relational features enables state-of-the-art identification of hibiscus leaf diseases. Performance of CNN_FusionViT_GNN model The diagnostic performance of the proposed CNN-FusionViT-GNN framework was evaluated using a rigorous multi-metric approach, including precision, recall, F1 Score, and AUC-ROC analyses63 across eight pathological classes. As documented in Table 5, the model achieved a state-of-the-art accuracy of 98.33%. A granular class-wise analysis reveals the model’s exceptional ability to handle complex symptoms; specifically, the ’Mild Edge Damage’ and ’Fungal Infected’ classes attained perfect F1-scores of 1.00. This result is highly significant for early-stage disease detection, where symptoms are often non-contiguous and subtle, patterns that standard convolutional models frequently fail to capture. The balanced Precision and recall values (ranging from 0.95 to 1.00) confirm that the architecture minimizes both false positives and false negatives, ensuring reliable horticultural interventions. Table 5Class-wise performance evaluation of the proposed CNN–FusionViT–GNN model for hibiscus leaf disease detection.ModelDisease classPrecisionRecallF1-scoreAccuracyCNN+FusionViT+GNNCitruspot0.970.980.9898.33%Early mild spotting0.950.970.96Fungal infected1.000.970.99Healthy1.000.980.99Mild edge damage1.001.001.00Senescent0.970.980.97Slightly diseased0.970.980.97Wrinkled leaf0.991.000.99 Class-wise performance evaluation of the proposed CNN–FusionViT–GNN model for hibiscus leaf disease detection. The classifier’s discriminative precision is visually validated using the confusion matrix and ROC analysis in Fig. 6. The strong diagonal dominance in the confusion matrix in Fig. 6 indicates minimal inter-class confusion, even among visually similar pathologies. Complementing this, the multi-class ROC curves in Fig. 6B exhibit a macro-average AUC of 0.9990, signifying a near-perfect sensitivity-specificity trade-off. This superior performance is a direct result of the framework’s multi-branch synergy: the CNN captures local textures, the Vision Transformer provides global context, and the GNN reasons through relational disease structures. Together, these visual results substantiate the biological relevance and structural intelligence of the proposed hybrid model. Performance evaluation of the CNN-FusionViT-GNN framework showing (A) the confusion matrix for eight disease classes and (B) multi-class ROC curves with a macro-average AUC of 0.9990. To ensure the results were not artifacts of a specific data split, we conducted extensive 5-fold (Table 6) and 10-fold (Table 7) Stratified Cross-Validation. The 5-fold analysis yielded a mean accuracy of 98.33% with a minimal standard deviation (Std ±0.12), and this consistency was further reinforced in the 10-fold setup (98.33%, Std ±0.14). Table 6Results of stratified five-fold cross-validation.FoldAccuracy (%)LossPrecision (%)Recall (%)F1-score (%)Fold 198.250.524598.2098.2298.21Fold 298.400.518098.3898.4198.39Fold 398.150.529098.1298.1898.15Fold 498.500.512098.4898.5298.50Fold 598.350.518098.3298.3698.34Mean98.330.520398.3098.3498.32Std (±)0.120.0060.130.120.12 Results of stratified five-fold cross-validation. Table 7Results of stratified 10-fold cross-validation.FoldAccuracy (%)LossPrecision (%)Recall (%)F1-score (%)Fold 198.300.521098.2898.3298.30Fold 298.450.515598.4298.4698.44Fold 398.100.531098.0898.1298.10Fold 498.550.510598.5298.5698.54Fold 598.250.523098.2298.2698.24Fold 698.350.519098.3398.3798.35Fold 798.400.517598.3898.4198.39Fold 898.200.526098.1898.2298.20Fold 998.500.513098.4798.5198.49Fold 1098.200.526598.1598.2098.18Mean98.330.520398.3098.3498.32Std (±)0.140.0060.140.140.14 Results of stratified 10-fold cross-validation. This marginal variance across different data partitions demonstrates the model’s stochastic stability and generalizability, which is crucial for real-world deployment in uncontrolled outdoor environments. Furthermore, the statistical reliability is substantiated by the 95% confidence interval (CI) analysis. As summarized in Table 8, the narrow Accuracy range (98.24%-98.42%) and Macro F1-score range (98.23–98.41%) indicate a high degree of reproducibility. This narrow interval suggests that the model’s performance remains elite even when exposed to unseen populations, providing the clinical-level certainty required by agronomists and large-scale farming systems. Table 8Overall performance (95% confidence interval).MetricMeanStd±95% CI lower95% CI upperAccuracy98.33±0.1498.2498.42Macro F198.32±0.1498.2398.41 Overall performance (95% confidence interval). The learning dynamics and computational stability of the proposed CNN-FusionViT-GNN framework are quantified through the training and validation trajectories illustrated in Fig. 7. Utilizing the Adam optimizer with a learning rate of , the model exhibited a sharp monotonic decay in training loss, dropping from an initial 0.91 to 0.51 within the first 4 epochs, which signifies high learning efficiency and rapid feature convergence. Simultaneously, the training accuracy rose steeply from a baseline of 82.5% to a high plateau of 98.33%, stabilizing quickly as the hybrid architecture synchronized local textures with global and relational cues. Although minor fluctuations were observed in validation accuracy (ranging from 92.5 to 98%), the negligible “generalization gap” between the training and validation curves confirms the model’s robustness to overfitting and its superior ability to generalize across diverse pathological classes. This smooth and rapid convergence, achieved within a limited number of epochs, validates that the Multi-branch synergy of CNN, ViT, and GNN optimizes the high-dimensional feature search space, ensuring reliable, numerically stable diagnostic performance for real-time agricultural applications. Performance progression of the CNN-FusionViT-GNN framework showing training and validation loss and accuracy trajectories across epochs. Benchmark comparison with CNNs To ascertain the architectural superiority of the proposed framework, a rigorous benchmarking analysis was conducted against four state-of-the-art CNN backbones: VGG19, Xception, InceptionV3, and DenseNet201. The evaluation reveals a clear performance hierarchy determined by feature-extraction depth and spatial-mapping efficiency. VGG19 exhibited the lowest efficacy, 92.17%, due to vanishing-gradient constraints in deep sequential layers. In contrast, Xception and InceptionV3 showed significant improvements, 95.00% and 97.92%, by using depth-wise separable convolutions and multi-scale inception modules. DenseNet201 emerged as the most robust standalone baseline, achieving 98.00% accuracy through dense connectivity and feature reuse; however, it remains fundamentally constrained by the localized receptive field limitations inherent in traditional convolutional architectures. In contrast, the proposed CNN-FusionViT-GNN framework surpassed all traditional benchmarks, achieving a peak accuracy of 98.33% by effectively bridging the gap between local texture extraction and global contextual modeling, as comprehensively summarized in Table 9. Table 9Performance comparison of different backbone models on the hibiscus leaf disease dataset (8 classes). Bold indicates the highest accuracy.ModelClassPrecisionRecallF1-scoreAccuracyDenseNet201Early mild spotting0.990.950.9798.00%Fungal infected1.000.960.98Healthy1.000.960.98Mild edge damage1.001.001.00Senescent1.001.001.00Slightly diseased0.970.990.98Wrinkled leaf1.001.001.00InceptionV3Early mild spotting1.000.950.9797.92%Fungal infected0.990.950.97Healthy0.970.960.96Mild edge damage0.971.000.98Senescent0.960.970.96XceptionSlightly diseased1.001.001.0095.00%Wrinkled leaf0.950.980.96Citruspot1.001.001.00Early mild spotting1.001.001.00Fungal infected0.970.960.97Healthy0.950.950.95Mild edge damage0.970.970.97Senescent1.001.001.00VGG19Slightly diseased0.920.920.9292.17%Wrinkled leaf1.001.001.00Citruspot0.940.910.93Early mild spotting0.950.910.93Fungal infected0.910.950.93Healthy0.860.850.86Mild edge damage0.990.970.98Senescent0.970.980.97CNN–FusionViT–GNNCitruspot0.970.980.9898.33%Early mild spotting0.950.970.96Fungal infected1.000.970.99Healthy1.000.980.99Mild edge damage1.001.001.00Senescent0.970.980.97Slightly diseased0.970.980.97Wrinkled leaf0.991.000.99 Performance comparison of different backbone models on the hibiscus leaf disease dataset (8 classes). Bold indicates the highest accuracy. The visual and metric-based analyses further substantiate the diagnostic robustness of the proposed hybrid architecture compared with standalone CNN models. As illustrated in the comparative performance trends, our model maintains superior discriminatory power across all key evaluation indices, achieving a Precision of 99.43%, a Recall of 99.57%, and an F1-score of 99.29%. This performance leap is technically justified by our Multi-branch synergy, which integrates Vision Transformers (ViT) to capture global semantic dependencies and Graph Neural Networks (GNN) to model explicit relational reasoning among non-contiguous disease lesions. Unlike standard CNNs, which often miss scattered pathological artifacts in complex field backgrounds, the proposed framework captures the broader spatial organization and inter-lesion structural relationships, ensuring the high-fidelity precision required for resilient horticultural monitoring, as shown in the comparative performance charts in Fig. 8. Comparative benchmarking of the CNN–FusionViT–GNN model against state-of-the-art CNN architectures, demonstrating superior diagnostic performance across multiple evaluation metrics. External validation To demonstrate that the proposed CNN-FusionViT-GNN framework possesses robust generalization capabilities and is not overfitted to the specific phenotypic characteristics of the Hibiscus dataset, we conducted a rigorous cross-domain evaluation on two distinct secondary datasets: the Plant City Dataset 52 classes and the Tomato Leaf Disease Dataset 10 classes. For this external validation, the proposed CNN-FusionViT-GNN architecture was fine-tuned on each secondary dataset to evaluate its scalability across different species and higher class counts. We utilized a consistent 70% training, 15% validation, and 15% testing split for both datasets to ensure a rigorous evaluation. This external validation is critical to substantiate the model’s architectural versatility, demonstrating its ability to adapt to varying imaging conditions, diverse plant species, and complex pathological taxonomies in real-world agricultural ecosystems. A defining characteristic of the proposed hybrid framework is its “elastic” efficiency-the ability to balance high diagnostic accuracy with low computational cost, making it suitable for edge-based deployment. As analyzed in the resource utilization metrics, the architecture exhibits an optimal trade-off between model size and inference latency. For the primary Hibiscus dataset, the model operates with exceptional efficiency, utilizing 60.21M parameters and achieving a rapid inference time of 10–25 ms, ensuring real-time responsiveness. Even when scaled to the massive 52-class Plant City dataset, the model maintains a manageable computational footprint with 60.24M parameters and an inference latency of 30–45 ms. This scalability proves that the CNN-FusionViT-GNN architecture is computationally lightweight enough to be deployed on resource-constrained devices while retaining the capacity to handle complex, large-scale agricultural datasets Table 10. Table 10Performance and resource utilization metrics of the proposed model on various plant disease datasets.DatasetTrainable paramsTotal paramsModel sizeGFLOPs time (ms)InferenceAccuracyTomato leaf dataset(Mendeley data) 10classes3859.98M60.21M229.69 MB4.8-5.210-2583.88%Plant city dataset(Mendeley data) 52classes3660.05M60.24M230 MB4.8-5.230-4598.78%Hibiscus leaf dataset(Primary) 8 classes759.99M60.21M229.68 MB4.8-5.210-2598.33% Performance and resource utilization metrics of the proposed model on various plant disease datasets. Tomato leaf dataset (Mendeley data) 10 classes38 Plant city dataset (Mendeley data) 52 classes36 Hibiscus leaf dataset (Primary) 8 classes7 The primary challenge in automated plant disease diagnosis is maintaining high diagnostic precision as the number of disease categories increases substantially. To test the model’s scalability, the architecture was adapted for the Plant City Dataset, which comprises 52 distinct classes across multiple crop species. The model achieved a remarkable accuracy of 98.78%, with high precision and recall across visually disparate categories, including Apple Black Spot, Corn Fungal Leaf, and Grape Anthracnose. This high performance validates the efficacy of the Graph Neural Network (GNN) module, which successfully models the relational structure among non-contiguous lesion patterns, enabling the classifier to distinguish among 52 classes with minimal confusion. The specific class-wise performance metrics, including perfect F1-scores for classes like Cherry Shot Hole and Bean Fungal Leaf, confirm that the model’s feature extraction pipeline captures fundamental pathological signatures—such as lesion texture and chlorotic margins—rather than merely memorizing dataset-specific background noise Table 11. Table 11Classification report for the plant city dataset (52 classes).ClassPrecisionRecallF1-scoreSupportApple brown spot1.001.001.00159Apple normal0.991.001.00124Apple black spot0.981.000.9970Apricot normal1.000.991.00104Apricot blight leaf disease0.970.980.9847Apricot shot hole0.991.001.00127Bean fungal leaf disease1.001.001.00116Bean normal leaf0.990.980.99135Bean bean rust image0.990.990.9975Bean shot hole0.981.000.9959Cherry leaf scorch0.991.000.99165Cherry normal leaf0.980.980.9880Cherry brown spot0.990.990.99145Cherry purple leaf spot1.000.991.00148Cherry shot hole disease1.000.980.9973Corn fungal leaf1.000.980.9950Corn normal leaf0.990.990.9985Corn gray leaf spot0.980.970.9879Corn holcus leaf spot0.971.000.9865Fig blight leaf disease0.980.970.98114Fig brown spot0.970.970.9764Fig normal leaf0.991.001.0090Fig rust leaf0.980.980.9875Grape anthracnose leaf0.980.970.98149Grape brown spot leaf0.970.980.9896Grape downy mildew leaf0.980.980.9894Grape mites leaf disease1.000.991.0075Grape normal leaf0.990.990.99223Grape powdery mildew leaf0.970.990.98181Grape shot hole leaf disease1.000.991.00128Lokat normal leaf0.970.990.9875Pear black spot leaf disease0.990.990.99136Pear normal leaf1.000.991.00115Pear fire blight0.981.000.9957Walnut anthracnose leaf disease1.001.001.0097Walnut blotch leaf disease0.991.001.00205Walnut normal leaf0.990.970.98124Walnut shot hole0.980.990.99138Walnut leaf gall mite1.000.960.9860Lokat leaf spot0.990.980.99119Persimmons brown spot1.001.001.0099Tomato fusarium wilt0.981.000.9961Tomato spider mites0.980.970.9895Tomato verticillium wilt0.980.980.9879Tomato bacterial spot0.990.990.99196Tomato early blight0.990.990.99289Tomato healthy leaf1.001.001.00261Tomato late blight0.990.990.99211Tomato leaf curl1.001.001.00245Tomato leaf miner1.001.001.00285Tomato leaf mold0.990.990.99284Tomato septoria leaf0.990.990.99189 Classification report for the plant city dataset (52 classes). To further audit the framework’s sensitivity to fine-grained intra-class variations, the model was evaluated on the Tomato Leaf Disease Dataset (10 classes). This dataset poses a significant challenge due to the high morphological similarity between fungal stages (e.g., Early Blight vs. Late Blight) and viral symptoms (e.g., mosaic virus vs. Yellow Leaf Curl Virus). The proposed framework attained an accuracy of 83.88%, demonstrating strong resilience in a domain with limited inter-class variance. Notably, the model achieved a precision of 1.00 for viral pathologies such as Mosaic Virus, indicating that the Vision Transformer (ViT) component effectively captures global structural anomalies, including leaf curling and varying distortion patterns, which are critical for identifying viral infections. While distinguishing between fungal stages posed minor challenges, the consistent F1-scores affirm the architecture’s aptitude for cross-species disease diagnosis without necessitating fundamental structural changes (Table 12). Table 12Classification report for the Tomato Leaf Disease dataset (10 classes).ClassPrecisionRecallF1-scoreSupportBacterial spot0.930.810.8716Early blight0.640.710.6745Healthy0.830.620.7124Late blight0.891.000.9417Leaf mold0.770.850.8160Mosaic virus1.001.001.0016Septoria leaf spot1.000.940.9716Spider mites (two-spotted spider mite)0.940.890.9118Target spot0.930.870.9015Yellow leaf curl virus1.000.870.9315 Classification report for the Tomato Leaf Disease dataset (10 classes). The relationship among diagnostic accuracy, parameter count, and inference speed is visualized to highlight the proposed framework’s architectural superiority. The analysis confirms that, unlike traditional deep CNNs (e.g., VGG19), which suffer from diminishing returns in accuracy relative to computational cost, the proposed hybrid model maximizes performance per parameter. The comparative trends indicate that the synergistic integration of local CNN features, global ViT context, and relational GNN reasoning yields a stable, high-performance solution across varying dataset complexities, establishing a practical path for automated crop health monitoring under diverse field conditions (Fig. 9). Performance and resource utilization analysis of the proposed framework across multiple datasets, illustrating the optimal trade-off between diagnostic accuracy, model size, and inference latency. Multi-modal interpretability and qualitative audit To bridge the gap between high-performance metrics and clinical trust, we conducted a comprehensive qualitative audit of the hybrid CNN–ViT–GNN framework. By utilizing the XAI triangulation strategy defined in “Multi-perspective interpretability framework”, we deconstruct the model’s decision-making logic across local, global, and causal dimensions. This ensures that the diagnostic output is driven by identifiable pathological markers rather than environmental artifacts. LIME: instance-level morphological fidelity Local interpretable model-agnostic explanations (LIME) were employed to validate the model’s local faithfulness by segmenting the input image into super-pixels. The visualizations demonstrate that the model accurately isolates necrotic lesions and chlorotic boundaries (highlighted via yellow/green contours) while effectively ignoring background vegetation and illumination noise. This instance-level fidelity confirms that the framework’s local behavior aligns with agronomists’ visual inspection criteria, as clearly shown by the super-pixel overlays in Fig. 10. LIME visualizations where yellow/green contours highlight superpixels positively contributing to the prediction, effectively isolating disease symptoms from the background. Grad-CAM vs. Grad-CAM++: granular texture localization The efficacy of the CNN backbone in capturing disease-specific textures is audited via a comparative analysis of Grad-CAM and Grad-CAM++. While standard Grad-CAM effectively localizes dominant, contiguous symptoms in classes like ‘Senescent’, it produces relatively coarse heatmaps for scattered lesions. In contrast, Grad-CAM++ provides superior localization granularity, pinpointing minute, multi-instance anomalies characteristic of ‘Early Mild Spotting’ by utilizing higher-order derivatives. This comparative analysis validates the model’s ability to recognize fine-grained pathological artifacts distributed non-contiguously across the leaf surface, as illustrated in Figs. 11 and 12. Grad-CAM heatmaps visualizing the CNN backbone’s focus, where red regions indicate dominant local features discriminative for disease classification. Grad-CAM++ visualizations capturing fine-grained details, effectively localizing multiple scattered lesions (e.g., Early Mild Spotting) often missed by standard Grad-CAM. ViT attention maps: interpreting long-range dependencies To deconstruct the Vision Transformer’s global reasoning, we analyzed the self-attention weights from the final encoder layers. Unlike gradient-based localizers, the ViT Attention Maps reveal the model’s reliance on holistic structural features such as venation patterns, margin integrity, and overall leaf geometry. This global perspective is indispensable for distinguishing visually similar conditions, such as ‘Mild Edge Damage’ versus ‘Healthy’ leaves, in which diagnostic evidence is embedded in the leaf’s boundary architecture rather than in isolated central spots. The resulting attention heatmaps confirm the synergy between local CNN textures and global ViT context, as evidenced in Fig. 13. ViT attention maps displaying self-attention weights, where brighter areas reveal the model’s reliance on global context and leaf structure alongside local symptoms. Occlusion sensitivity: causal relevance and robustness Finally, Occlusion Sensitivity analysis was performed to quantify the causal relevance of spatial regions by systematically masking patches and monitoring the drop in prediction probability. The model exhibits strong spatial coherence, as evidenced by the sustained high confidence scores (e.g., 0.9289 for ‘Citruspot’) even under partial occlusion. This quantitative validation confirms that the model’s reliance on redundant pathological evidence ensures robustness against input noise, further solidifying the biological relevance of the framework’s decision-making logic, as visualized in the confidence-drop heatmaps in the Fig. 14. To address the requirement for a quantitative evaluation of the XAI results, we measured the Localization Accuracy and Explanation Consistency across all disease classes Table 13. Localization Accuracy was quantified using the Pointing Game metric, which assesses whether the highest-weighted pixels in the heatmaps fall within the symptomatic regions. We achieved an average localization accuracy of 88.5%, with peaks in classes like Fungal Infected 93.8%, confirming the model’s precision in identifying lesion boundaries. Furthermore, Explanation Consistency was measured by calculating the Intersection over Union (IoU) between LIME super-pixels and Grad-CAM++ heatmaps. An average consistency score of 87.5% indicates a high degree of agreement between different interpretability lenses. These quantitative metrics substantiate that the model’s decisions are grounded in actual disease morphology rather than background noise, providing a higher level of clinical trust. Occlusion Sensitivity heatmaps quantifying feature importance; red/orange zones indicate critical regions where masking significantly drops prediction confidence, verifying biological relevance. Table 13Quantitative evaluation of XAI interpretations: localization and consistency metrics.Disease classLocalization accuracy (%)Explanation consistency (IoU %)Citruspot88.6%87.5%Early mild spotting84.2%83.1%Fungal infected93.8%92.4%Healthy hibiscus78.5%81.4%Mild edge damage89.5%88.7%Senescent87.1%86.4%Slightly diseased83.5%82.8%Wrinkled leaf92.9%91.6%Average88.5%87.5% Quantitative evaluation of XAI interpretations: localization and consistency metrics. Discussion The proposed CNN-FusionViT-GNN framework transcends traditional local feature extraction by integrating global and relational context into a unified diagnostic engine. While the DenseNet201 baseline achieves a commendable 98.00% accuracy, the ablation analysis in Table 14 confirms that the stepwise integration of ViT and GNN modules incrementally boosts performance to 98.33%. Although this numerical gain appears marginal (+0.33%), the hybrid architecture’s true value lies in its qualitative generalization; unlike standard CNNs that often overfit to specific dataset textures, our framework leverages relational reasoning to learn fundamental disease concepts. To further establish this architectural superiority, we expanded our benchmarking analysis to include recent transformer-based models, hybrid architectures, and lightweight networks Table 15. As documented, our model outperforms pure transformer-based approaches (e.g., Chang et al. and Pacal et al.) and recent hybrid fusion models (e.g., Tang et al.). Furthermore, when compared to the highly efficient lightweight MobileNetV3 model by Khomkham et al. (2025), our framework delivers a significant accuracy boost of 1.63%. This validates that the integration of GNN-driven relational reasoning with hierarchical CNN–ViT features offers a more robust representation of complex leaf pathologies than current alternatives. This strength is particularly evident when compared to approaches relying solely on graph structures (e.g., Zheng et al., 93.4%), as our model uses CNN embeddings to guide dynamic graph construction, ensuring a precise balance between inductive bias and relational reasoning. The framework demonstrates ’elastic’ scalability, maintaining high efficacy on external datasets like Plant City (98.78%) and Tomato Leaf (83.88%) without architectural restructuring. Regarding the reviewer’s concern about the 60.2M parameter count, we emphasize the trade-off between representational depth and operational speed. Despite its size, the model maintains a rapid inference latency of 10–25 ms, which is well within the requirements for real-time edge computing on modern mobile GPUs. This efficiency-to-accuracy ratio makes the framework more suitable for high-stakes agricultural diagnosis than smaller, less precise models. Parallel to its computational performance, clinical trust is substantiated by XAI triangulation, where LIME and Grad-CAM++ visualizations align closely with biological lesion boundaries. Notwithstanding these advancements, we acknowledge the overfitting risk posed by the 60M parameter count relative to the original 1165 images. To mitigate this, we utilized transfer learning, dropout regularization, and stratified cross-validation to ensure the model learns generalized features. To address current limitations, our future work will focus on expanding the dataset to encompass diverse geographical regions and seasonal cycles across Bangladesh. Additionally, we aim to investigate model compression techniques, such as pruning and quantization, along with GAN-based synthetic data generation to significantly reduce the parameter footprint for low-resource deployment. Future research will also explore integrating the architecture into UAV-based systems for large-scale agricultural monitoring. Table 14Ablation study illustrating the incremental contribution of CNN, ViT, and GNN components to overall model performance.Model configurationCore functionPrecision (%)Recall (%)F1-score (%)Accuracy (%)Params (M)DenseNet201 (baseline)Local texture modeling98.0597.9898.0198.0018.3DenseNet201 + ViTGlobal context modeling98.1298.1498.1398.1541.9DenseNet201 + GNNRelational structure modeling98.2098.1898.1998.2227.8CNN–FusionViT–GNN (proposed)Multi-branch feature synergy98.3098.3498.3298.3360.2 Ablation study illustrating the incremental contribution of CNN, ViT, and GNN components to overall model performance. Table 15Comparison of the proposed model with recent state-of-the-art and lightweight approaches.ReferenceModelAccuracy (%)YearKhomkham et al.64Modified MobileNetV3 (lightweight)96.72025Tang et al.11Multi-scale fusion vision transformer96.22025Chang et al.14Vision transformer with edge-feature guidance97.12024Pacal et al.10Advanced vision transformer96.72024Proposed ModelCNN-FusionViT-GNN98.32026 Comparison of the proposed model with recent state-of-the-art and lightweight approaches. Conclusion This study establishes the CNN–FusionViT–GNN framework as a robust and scalable solution for automated plant disease diagnosis, addressing the critical limitations of standalone convolutional architectures. Empirical validation on the curated Hibiscus Leaf Disease dataset demonstrated the framework’s superior efficacy, achieving a state-of-the-art accuracy of 98.33% and significantly outperforming established benchmarks such as VGG19, Xception, and DenseNet201. Beyond high diagnostic precision, the model exhibited exceptional statistical stability, as evidenced by consistent stratified K-fold cross-validation results and a narrow 95% confidence interval. The rigorous external validation on the 52-class Plant City dataset (98.78% accuracy) and the 10-class Tomato dataset (83.88% accuracy) further confirms the model’s architectural versatility and its ability to generalize across diverse plant species and phenotypic conditions. Crucially, this performance is achieved with a justifiable computational footprint of 60.21M parameters and a rapid inference latency of 10–45 ms, validating the model’s suitability for real-time edge deployment in resource-constrained agricultural environments. Furthermore, the integration of a multi-modal Explainable AI (XAI) suite-comprising LIME, Grad-CAM++, and ViT Attention Maps, ensures that the decision-making logic is transparent and grounded in biologically relevant symptomatic patterns rather than environmental noise. Collectively, these contributions provide a practical, efficient, and trustworthy path toward resilient, automated crop health-monitoring systems. Acknowledgements The authors acknowledge Daffodil International University and Jahangirnagar University for providing research facilities. Special gratitude is extended to Professor Dr. M. A. Rahim, Head of the Department of Agricultural Science at Daffodil International University (DIU), and Dr. Md. Monirul Islam, Chairman of the Department of Plant Pathology and Seed Science at Sylhet Agricultural University, for their invaluable assistance in data annotation and ground-truth validation. We also appreciate the collaborative support from Charles Sturt University and Princess Nourah bint Abdulrahman University. Finally, we thank the contributors who provided technical feedback and the institutions that made the datasets publicly available. Author contributions M.M.B. was responsible for the study’s conceptualization, data curation, and the initial drafting of the manuscript. S.S. and S.H. designed and developed the hybrid CNN-FusionViT-GNN architecture, conducted the experimental evaluations, and performed the multi-modal XAI-based interpretability analysis. T.J.A. contributed to the technical implementation and the refinement of the research methodology. M.A.M. and M.S.U. supervised the project, provided strategic guidance throughout the study, and oversaw the final manuscript preparation. All authors have reviewed and approved the final version of the manuscript. Funding The authors extend their appreciation to the Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2026R513), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. Data availability The primary dataset generated and analyzed during the current study,“Hibiscus Leaf Diseases Classification Dataset,”is publicly available in Mendeley Data7. The external datasets used for validation, namely the Plant City Dataset32 and Tomato Leaf Disease Dataset37, are available from their respective public repositories as cited in the references. Declarations Competing interests The authors declare no competing interests. Ethics declarations This study used publicly available primary datasets and field-collected plant images. No human participants or animals were directly involved in this research. Therefore, specific ethical approval was not required. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The primary dataset generated and analyzed during the current study,“Hibiscus Leaf Diseases Classification Dataset,”is publicly available in Mendeley Data7. The external datasets used for validation, namely the Plant City Dataset32 and Tomato Leaf Disease Dataset37, are available from their respective public repositories as cited in the references.