Tomatoes (Solanum lycopersicum) is one of the most widely planted vegetable crops worldwide, with great significance in food security, economics and nutrition. However, tomato cultivation is severely affected by several bacterial, fungal and viral diseases namely bacterial spot, early blight, late blight, leaf mold, septoria leaf spot and mosaic virus. These diseases are detrimental to yield and fruit quality and negatively impacts income of farmers.

Traditionally, diagnosis of plant diseases is made based on visually observing leaves/fruits using inspection of agricultural experts. While the expert-based method is successful it is time consuming, subject to human judgment,laborious, expensive and difficult to apply to a large agricultural field. Since then, there has been immense interest in using Artificial Intelligence (AI) based deep learning models for automated diagnosis of plant diseases.

Deep learning architectures namely CNNs, transfer learning, transformer based and ensemble models have reported good performance in tomato leaf disease classification. Models such as VGG16, ResNet50, DenseNet121, EfficientNet and Vision Transformers are successfully used for lesion recognition in recent researches. Despite such progress, some research challenges still persist in automated diagnosis of tomato leaves:
 
Research gap
 
Most existing tomato disease classification approaches rely on end to end CNN pipelines, focusing more on global image feature representations while neglecting local and lesion centric learning. Furthermore,
1. Background clutter, noisy irrelevant regions tend to result in classification inaccuracies.
2. Similarity between classes (spider mites, target spot, leaf mold, etc.) leads to classification confusion.
3. Few methods attempt to explicitly model disease to disease relationship.
4. Context aware graph-based correlation learning has rarely been utilized for tomato disease classification.
5. Hybrid decision fusion approaches using prototype learning and graph similarity scoring are rarely explored.

The objective of this paper is to address these shortcomings by introducing a novel hybrid deep learning framework, incorporating APD (Adaptive Patch Decomposition), GDCL (Graph Based Disease Correlation Learning) and HDFC (Hybrid Decision Fusion Classification). This proposed framework addresses the need for a context-aware local-global feature extraction model by using both lesion-centric feature learning with graph-based inter-class relationship learning and hybrid decision fusion. The contributions are as follows:
● Designing an adaptive patch decomposition based on lesion-centric feature extraction.
● Presenting a graph-based disease correlation learning in inter-class disease similarity.
●  Integrating local and global contextual feature aggregation.
● Presenting a hybrid decision fusion classifier combined with prototype distance and graph similarity learning.
●  Comprehensive experiments and comparative evaluation through 5-fold cross-validation.
 
Literature review
 
The automatic tomato leaf disease identification has become a hot topic and attracts more attention in the recent years for the necessity of precision agriculture and the rapid progress of early crop disease detection. Many researchers propose the deep learning approaches to increase the recognition performance. Saeed et al., (2023) proposes a transfer learning-based CNN framework for smart tomato leaf disease detection. The paper shows that pre-trained CNN architecture improve the classification performance greatly but decrease the computation complexity and increase speed compared to existing approaches. Khan et al., (2023) proposes TomFormer which is a novel hybrid transformer-CNN structure for premature and accurate identification of tomato diseases and it is strong enough to handle the classification problem across datasets and indicate the crucial role of attention based global feature learning. Ouamane et al., (2024) proposes a CNN based framework in combination of tensor subspace learning and HOWSVD-MD, reporting that the classification precision and robustness of tomato leaf diseases are increased. Sharma et al., (2025) proposes an ensemble framework combining ResNet50 and MobileNetV2 in order to achieve excellent classification performance and gain more feature diversity. Sujatha et al., (2025) demonstrates an integrated ML and DL framework for leaf infection finding and claims the robustness of the system relies on both preprocessing and hybrid architecture. Das et al., (2025) provides the latest literature review on detection, classification and segmentation for tomato leaf diseases utilizing deep learning algorithms including CNN, transformers and object detection models. Ramos and Sappa (2025) analyses YOLO architectures on tomato leaf disease identification in detail and proves that object detection-based approach plays important role on locating the leaf lesion. Gunasekaran et al., (2026) concentrates on the application of trivial deep learning and fringe AI based approach for real-time tomato disease identification and pays attention to the challenges for deploying models into mobile and IoT devices.

In their research of plant leaf infection finding with deep learning methods, Zhao et al.  (2025) stated that CNN based methods are widely used while the transformer-based methods are appearing frequently in recently published articles. Mustofa et al. (2023) delivered an in-extensive assessment of deep learning standards like CNN, ViT, YOLO and residual networks for plant disease detection. Hosen and Islam (2025) examined VGG19 and model InceptionV3 for leaf disease classification in tomato plant and achieved very promising result to real-world crop health inspection. Previous researches also applied several classical CNN models including VGG16, ResNet50, DenseNet121 and EfficientNet with very well performance, however the issues with inter-class similarity and learning contextual features are often observed. Nevertheless, researches on graph-based disease correlation learning, patch-wise adaptive feature decomposition and hybrid score fusion are seldom.

Metagar and Walikar (2024) in the review article, published in Agricultural Science Digest, outlined machine learning (ML) and deep learning (DL) techniques employed in the diagnosis of plant diseases and noted that image processing techniques, feature extraction and the combination of sophisticated learning models contribute to higher accuracy in disease prediction, thereby enabling precision agriculture practices. Mehtt et al., (2025) through their article presented in the Indian Journal of Agricultural Research provided an overview of different AI techniques to aid in disease diagnosis of various crops and concluded that convolutional neural networks (CNN), transfer learning and other deep learning frameworks have shown to be a robust and scalable way to perform plant disease classification, thereby powering smart farming operations.

Mehtt et al., (2025) also conducted another review article in the Indian Journal of Agricultural Research, analysed the recent research related to the application of artificial intelligence for detecting plant diseases where they observed that DL algorithms integrated with computer vision for plant images have made improvements in classification and have been proved useful for timely detection of plant diseases in various agricultural setups. Nigam and Jain (2020) analysed a plethora of deep learning algorithms used in the detection and identification of plant diseases in their review, concluded that there have been immense improvements in the classification accuracies using the applications of computer vision (CV), machine learning (ML) and deep learning (DL) for intelligent decision-making process in agriculture.

Most of existing work follows an end-to-end CNN pipeline and implicitly extracts features without modeling inter-class disease relations. This paper attempt to tackle this problem through combining the APD, GDCL and HDFC together.

Data description
 
The designed system was tested on a tomato leaf diseases dataset consisting of 20000 labeled images acquired from publicly available datasets on PlantVillage and agriculturally enriched dataset. The classes considered in this dataset are ten:
1. Bacterial spot.
2. Early blight.
3. Late blight.
4. Leaf mold.
5. Septoria leaf spot.
6. Spider mites.
7. Target spot.
8. Yellow leaf curl virus.
9. Mosaic virus.
10. Healthy leaves.
The dataset was divided into:
● 70% Training set.
● 15% Validation set.
● 15% Testing set.

Data augmentation was performed to enhance the generalization ability on the given data; rotation, flipping, zooming, brightness adaptation and random cropping were employed.
 
Implementation details
 
Proposed framework was implemented on Python with TensorFlow and PyTorch frameworks. Training was performed on NVIDIA RTX GPU architectures using following hyperparameters in Table 1.


 
Proposed approach
 
The designed hybrid deep learning structure for categorization of tomato leaf diseases will be described in this section. Our methodology consists of seven steps: adaptive preprocessing, adaptive patch decomposition, local feature learning, global context aggregation, graph-based disease correlation learning, feature re-weighting and hybrid decision fusion classification as shown in Fig 1.


  
The proposed framework can accurately classify the diseases by simultaneously utilizing spatial, context and inter-class correlation features.
 
Phase 1: Data acquisition and adaptive preprocessing
 
This research used 20,000 tomato leaf images labeled with their disease types. They can be formulated as:
 

 

Where,
Ii= Input image.
yi= The related disease label.
N = 20.000.

All the images are processed by adaptive preprocessing technique to maintain consistence and improve image quality as shown in Fig 2.


 
Image resizing
 
Each image size is modified by using the below formula:
 

 
This default size has made for fast processing with CNNs.
 
Noise filtering
 
Noise is deleted using adaptive filtering:

 

 
Where,
 = The filtering function.
 
Contrast enhancement
 
The sorted (filtered) image is enhanced the contrast by using the below formula:
 

 
Where,
 = The contrast enrichment operator.
 
Normalization
 
The image normalization can be done by using the below formula, i.e. normalized into the range [0,1]:
 

 

 
This method enhances the convergence of the model and training stability.
 
Phase 2: Adaptive patch decomposition (APD)
 
The pre-processed image is segregated into numerous disease-related patches:
 

 
Where,
p_i= The i^th patch.
Each patch can be denoted as:

 
Threshold segmentation method is used to remove background and irrelevant patches:
 

 
Where,
= A disease relevance pointer.

This phase confirms that only lesion-related sections are conserved for feature learning.
 
Phase 3: Local feature learning
 
To extract local features deeply, every single disease-relevant patch is passed through CNN backbone:
 

 
Where,
f_i = The feature vector of patch p_i.

The complete patch-level feature set is denoted as:
 

 
Each feature vector fits to:

 
Where,
d= The dimension of feature.

This phase depicts local lesion patterns such as “spots, blights, discoloration and texture irregularities”.
 
Phase 4: Global context aggregation
 
Patch-level features are grouped into global interpretation for preserving the characteristics of entire leaf disease:
 

 
Using average pooling as below:
 

Instead, attention-based weighted aggregation is extracted as:
 

 

 
subject to:
 

               

This phase describes global leaf morphology and disease diffusion.
 
Phase 5: Graph-based disease correlation learning (GDCL)
 
Graph structure is designed to model the relationships of inter-class diseases:
 

 
Where,
V= Disease classes and  indicates class relationships.

For representing ten disease classes in tomato leaf:
 

 

Similarity matrix
 
Inter-class similarity is processed using cosine comparison:
 

 

 
Where, 
S_ij =indicates resemblance between I and j classes.
Thus,
 

 
Graph transformation
 
Graph transformation is used to improvise the global features:
 

Where,
A= Adjacency matrix.
W= Trainable weight matrix .
= Activation function.
This phase enhances the discrimination of available features by disease correlation method.
 
Phase 6: Feature Re-weighting
 
Learnable attention method allots significant weights to refined features of tomato leaves.
 

Where,
 = Weight vector.
 = Element-wise multiplication

This highlights disease-specific discriminatory features.
 
Phase 7: Hybrid decision fusion classifier (HDFC)
 
The final classification is acquired by merging prototype distance and graph resemblance.
 
Prototype distance score
 

 
Where,
C_k= The centroid of class k.
 
Graph similarity score
 

 

Score fusion
 

 
subject to:
 

 
Final prediction
 

 
This merging upgrades stoutness and category consistency. The recommended structure is assessed using the metrics of accuracy, precision, recall and F1 score. This model achieved 99.03% accuracy, demonstrating superior performance for leaf disease classification in tomato plant.

The robustness and generality of proposed framework are evaluated by performing 5-fold cross-validation. The dataset was divided into 5 randomly chosen, equal-sized partitions. Four of the partitions were selected for the training process and one was selected for testing on each fold.

Summary of performance averages across five folds as indicated in Table 2.



The similarity of the folds reflects the stability of the proposed method.Contrary to most of the above work which uses one train-test evaluation, five-fold cross-validation was performed for stability and less bias in the results, showing high generalization.

The performance analysis report in Table 3 showed that the suggested hybrid system out-performed those CNN and transfer-learning system. The combination of APD, GDCL and HDFC help to increase the abilities of lesion detection, context learning and disease discrimination. The APD module eliminated effects from the complicated background by targeting to only those regions related to diseases which helps to improve feature quality and classifier stability greatly. The GDCL module captured similarity among visually similar categories and therefore reduced inter-class confusion. Furthermore, the HDFC module boosted prediction stability by fusing graph similarity value and learning based on prototype-distance.



The relative performance in Table 3, demonstrates how the presented hybrid model outperforms previous deep learning architectures. The Basic CNN model showed the poorest performance and achieved 91.42% accuracy. The deep models like VGG16 and ResNet50, provided better performances as accuracy raised to 94.87 and 96.54%. Then, performance reached 97.21 and 97.84 accuracy using DenseNet121 and EfficientNet-B0 because they provided an efficient and effective way for feature propagation and parameter reuse. In the end, InceptionV3 showed 98.96% baseline accuracy with good multi-scale feature learning. The recommended hybrid model proven that the chosen dataset performed well with 99.03% accuracy, 98.98% precision, 98.94% recall and 0.053 loss. Such achievement benefited from combined adaptive patch decomposition, graph-based disease correlation learning and hybrid decision fusion classification which not only can optimize the local lesion representation but also improve inter-class disease discrimination. Furthermore, it can be observed that the values of loss decreased over time from 0.241 of basic CNN to 0.053 of proposed model.
 
Disease class wise performance measures
 
The proposed framework performance is further justified by considering performance per disease class. There are 10 classes of tomato leaves were used in the experiment for which disease-wise performance were measured in below Table 4.



The class-wise performance of the model was mostly high across all disease classes. The healthy class recorded the best F1-score of 0.97, which confirms the clear separation between the healthy and disease samples. The Mosaic Virus class also has a high F1-score of 0.95, which reveals high accuracy in the detection of the virus. The class-wise performance for Spider Mites and Target Spot was lower at 0.85, possibly attributed to the visually similar symptoms and similar features of lesion spread between these diseases.
 
Overall performance metrics
 
Overall metrics in Table 5, From the macro average, it’s observed that the model performs fairly in all classes. The weighted averages recall of 0.92 shows that the model continues to perform well inspite of class imbalances.



From the performance diagram, it can be seen that the performances are improved gradually from conventional CNN networks to the proposed hybrid approach. The baseline Basic CNN has very low performance 91.42% and deeper transfer learning networks like ResNet50, DenseNet121 and InceptionV3 yield much higher accuracy on disease recognition. The proposed hybrid framework shows superiority among all the models owing to adaptive lesion-centric patch extraction, local and global feature combination, graph-based learning on disease relationships, hybrid fusion for decision making and reduced loss values 0.053 compared with 0.241 that indicates better convergence and generalization performance.
 
Analysis discussion
 
The overall high performance can be explained as a combination of the power of local lesion feature extraction and the strength of learning on global relationship among diseases. The APD module successfully isolates the regions containing disease to minimize the background influence. GDCL effectively learns the inter-class similarity for improved discrimination of visually similar diseases. Eventually, the HDFC aggregates both the distances of the prototypes from samples and the similarities in graphs for much robust prediction
 
Confusion matrix exploration
 
Class-wise, the exploration of confusion matrix in Fig 3 reveals the number of correctly classified and misclassified samples for each disease type of the suggested hybrid model. This indicates class-to-class confusion and delivers a signal of the consistency of the category model. Let the confusion matrix be represented as:
 

 
Where,
●C_ij= The number of samples belonging to class i predicted as class j.
●n= 10 disease classes.



The diagonal elements Cii represents the accurately predicted samples, while others in the off-diagonal elements represent wrongly predicted samples. The confusion matrix of the proposed hybrid model is having a very high diagonally dominant property.

The healthy class achieved perfect recall: Recall = 1.00  which means all healthy leaf images were correctly classified. Similarly, mosaic virus shows very high recall: Recall = 0.99 demonstrating excellent detection capability. Minor confusion is observed between: Spider mites, target spot, leaf mold. It’s also reasonable since those diseases often share similar textural pattern and discoloration appearance. The proposed APD + GDCL framework indeed reduces the inter-class overlaps, which is illustrated in the confusion matrix.
 
ROC curve analysis
 
The receiver operating characteristic (ROC) curve appraises the discriminative ability of the proposed model for each disease class.
The ROC curve plots:

 
against
 

 

 
Where,
TPR= True positive rate.
FPR= False positive rate.

The area under curve (AUC) is computed as:
 

 
Stronger classification performance is suggested by larger AUC value.

The proposed hybrid model is expected that the ROC curves are positioned around the top left corner for all ten disease classes. The estimated AUC values are summarized below in Table 4.

In Table 6, estimated AUC values, proven that Healthy and Mosaic Virus classes have highest AUC value which is related to highest precision and recall. Spider mites and Target spot have a moderately low AUC value that reflects class overlap.


 
Graphical interpretation
 
The ROC value typically demonstrates that the proposed model maintains high sensitivity with minimal false positives. This is mainly due to:
● Adaptive lesion patch extraction.
● Graph-based disease correlation refinement.
● Feature re-weighting.
● Hybrid score fusion.

Compared with baseline CNN models, the ROC curve of the proposed framework is significantly closer to the ideal classification boundary. The performance indexes, confusion matrix and ROC curves collectively show that the developed hybrid model has good and generalized performance of tomato leaf disease classification. APD contributes to better localization of lesion and GDCL effectively captures the correlations among diseases. Therefore, the overall classification performance is good.
 
Real-world applicability
 
The current framework can be utilized in:
● Smart farming systems.
● IoT enabled monitoring of agriculture.
● Mobile based disease diagnosis applications.
● Drone based crop monitoring.
● Precision agriculture decision support systems.

The processing of the lesion-centric patches efficiently at the hardware level due to lightweight patch processing mechanism is also useful for real-time application deployment.
 
Limitations
Though current model provided good results there are a few limitations:
1. The current dataset contains mostly lab controlled images and requires to be tested for varying field conditions and scenarios like illumination problems, occlusion etc.
2. For edge based real-time applications model optimization techniques are needed.
3. Factors like humidity, temperature and soil conditions are not considered in the current frame work.

A hybrid framework for automated tomato leaf disease classification based on an “adaptive preprocessing, Adaptive Patch Decomposition (APD), deep local feature learning, global context aggregation, Graph-Based Disease Correlation Learning (GDCL), feature re-weighting and a Hybrid Decision Fusion Classifier (HDFC)” is reported. This proposed method integrates both localized disease manifestations and holistic leaf traits, yielding higher classification performance. Through experiments, our method has achieved up to 99.03% accuracy, 98.98% precision, 98.94% recall and a lowest loss of 0.053 on the given dataset, surpassing many existing state-of-the-art deep learning methods. Hence, it is believed that this framework has high reliability for the purpose of early and precise plant disease diagnosis.

In addition, this hybrid method is also adapted for plant monitoring using mobile phones, Internet of Things (IoT) devices and drones and for integration with transformer-based attention and heterogeneous agricultural data, which are crucial for the new generation precision agriculture and smart crop health monitoring.
 

The present study was supported by the Departmentof Computer Science, Mother Teresa Women’s University, Kodaikanal, India. The authors sincerely express their heartfelt gratitude to all the faculty members of the Department of Computer Science for their valuable guidance, encouragement and continuous support throughout the completion of this research work.
 
Disclaimer
 
The views and conclusions presented in this article belong to the authors and may not represent the views of their affiliated institutions. The authors have made efforts to ensure that the information provided is accurate and complete. However, they do not take responsibility for any direct or indirect loss resulting from the use of this content.

The authors declare that there is no conflict of interest regarding the publication of this research article.