Tomatoes are the world’s second-most significant vegetable crop, but they are also affected by various biotic stressors, environmental variables and diseases. These diseases can spread quickly due to their contagious nature, with fungi being the primary source of infectious illnesses. Tomatoes commonly experience fungal diseases such as leaf spot, Cercospora leaf mould, Fusarium wilt, Grey leaf spot, Powdery mildew, Verticillium wilt, White mould, Alternaria stem canker, Corky root rot, Fusarium crown and root rot, Fusarium foot rot, Southern blight and Buckeye rot. Pests and diseases are the primary cause of agricultural productivity decline and crop losses, resulting in tonnes of crops lost annually (Selvia et al., 2014; Hasan et al., 2020).
       
To prevent significant losses and boost yields, early diagnosis of these illnesses is therefore essential. Conventional techniques for identifying plant diseases include the use of diagnostic specialists (Sabri et al., 2020), are costly in terms of time and effort, require specialised instruments, have a high diagnostic error rate and cannot be done with just the naked eye (Dhandapani et al., 2022). Recent advances in computer technology introduced a new system that can identify plant diseases by images to distinguish different diseases of plants at an early stage. This is a significant advancement in the field of Agrifood which can automate many processes in modern agriculture. The system utilises Artificial Intelligence methods to achieve desired outcomes. The usage of convolution neural networks is the basis of the model utilized as they are a high-performance deep learning network which is feasible for processes in tomato plants.
 
Related work
 
Numerous researchers are tackling the problem of disease of tomato leaf identification using various methods that combine image processing and machine learning to automate the process. Many academics have recently turned to deep learning algorithms in their hunt for more accurate results.
       
Machine learning was known as the most successful approach in detecting diseases in fish, plants and animals, according to some researchers (Cho et al., 2024; AlZubi, 2023; Koike, 2023; Wasik and Pattinson, 2024) and also in the manufacturing industry (Porwal, 2024). Kumar et al., (2019) utilized two PlantVillage datasets, specifically the 2019 ImageNet dataset, to examine the plots of “Visual Geometry Group (VGG) Net, LeNet, ResNet50 and Xception” to detect nine distinct tomato leaf diseases. Hasan et al., (2019) created a pre-trained network model that correctly recognised illnesses affecting tomato plants at a 94% to 55% accuracy level.
       
Another work adopted transfer learning where a neural network based on AlexNet as a deep learning framework was used for disease classification on tomato plant leaves; it achieved an accuracy rate of 95.75%, reported by Sangeetha and Rani (2021). The use of VGG 19 and AlexNet architectures helped in disease identification related to tomatoes at a striking speed of 13,262 frames per second. According to Rangarajan et al., (2018), the model was 97.49% correct. Kibriya et al., (2021) used two (CNNs), VGG16 and GoogLeNet, to classify tomato leaf diseases. This analysis’s chief objective is to identify the greatest method for recognizing diseases of tomato leaves by employing deep-learning techniques. Working with 10,735 leaf pictures from the Plant Village dataset, it was observed that VGG16 had an accuracy rate of 98.8%, while GoogLeNet fetched 99.23% accuracy, with the latter having a higher result.

Dataset collection
 
•    A dataset comprising images of tomato plants affected by leaf mold, early blight and late blight diseases was obtained from Mendeley.
•    The dataset consists of 4,362 images, including samples of healthy tomato plants for comparison (Fig 1).


 
Data preprocessing for CNN models
 
Data preprocessing for Convolutional Neural Network (CNN) input involves resizing images to 256×256 for consistency, normalizing pixel values, using data augmentation techniques like rotation, flipping and cropping to create training samples and incorporating additional information like labels or metadata to enhance input and model convergence. These steps significantly impact learning and inference processes, as well as the model’s ability to operate in various applications.
 
Architecture of convolutional neural network (cnn) model
 
Many components of CNN’s architectural paradigm promote learning efficiency. Convolutional layers are utilised to extract features, pooling layers to reduce features and softmax layers to classify them.
       
The hyperparameters are determined outside the algorithm and are established before training. There is no widely accepted methodology for determining the appropriate hyperparameters, leading to numerous experiments. Table 1 presents the hyperparameters utilised during the training of the model.


       
Convolution operations involve filters in convolutional layers, gradually covering the input picture or feature map from the preceding layer. Information is extracted from sections that overlap the filter at each stage. To get a value at a specific place on the feature map, the filter elements are multiplied by matching picture elements and the results are added. The convolution operation is defined over a two-dimensional kernel (K) and input image (I) as follows:

Where
i and j = Represent the image.
(I) = Coordinates.
m and n = Designate the kernel.
(K) = Coordinates.
       
The model uses pooling layers to reduce computational complexity and manage overfitting. The max pooling layer extracts the highest value from feature maps. The Fully Connected Layer then classifies the extracted features. Here, this layer was applied with Softmax function which predicts classes using numerical values based on information from the previous layer. The softmax function converts the numerical values χ1, χ2, χ3, … cn.of the neurons in the next layer into probabilities P1, P2, P3, …Pn. The probabilities are given as

Where
P_i = Probability of class.
i =  After applying softmax.
χ_i = The numerical value of the.
ith = Neuron in the preceding layer.
       
Next, the training model will be assessed with test data to determine how it performs with new data.
 
The evaluation matrix using classification parameters
 
The confusion matrix is a widely used approach for dealing with classification issues such as binary and multiclass classification. Accuracy, precision, recall and F1 score, commonly used measurement measures, were employed to analyse the proposed model’s performance.

The research developed Convolutional Neural Network (CNN) architectures to identify diseases in tomatoes by analyzing a dataset of healthy and diseased leaves. The CNN model achieved significant performance metrics, with a recorded loss of 0.1725 and an accuracy of 93.45% on the training dataset. It also achieved a validation loss of 0.2141 and an accuracy of 92.31% (Fig 2). Evaluated using an independent validation dataset, these measures demonstrate the model’s capacity to apply its learned patterns to new, unseen data. Although there is a small decline in accuracy compared to the training set, the validation performance remains impressive, indicating the model’s efficacy in detecting diseases in tomatoes through image analysis.


       
Further, the predictions and corresponding confidence scores for multiple instances are presented in Fig 3. The model consistently performs well across different classes. However, slight variations in accuracy and confidence scores offer useful insights into its ability to distinguish between each class.


       
The confusion matrix presents an analysis of the model’s classification accuracy for three specific categories: healthy leaf, leaf mold and early and late blight. Each of the rows in the matrix contains the actual labels for classes, whereas every column represents the expected class labels (Fig 4).


       
The CNN architectures have been evaluated for their effectiveness in detecting unhealthy leaves from various images, indicating their potential for disease diagnosis and plant healthcare. The model correctly detected both diseased and healthy leaves, indicating its capacity to distinguish between healthy and unhealthy tomato leaves.
       
The precision, recall and F1-score metrics are presented in Table 2. The model in this instance exhibits high accuracy in all classes, with precision ranging from 0.8137 to 1.0000, recall from 0.8646 to 0.9924 and F1-scores from 0.8384 to 0.9962. This demonstrates the model’s ability to accurately determine samples for each class. The model’s overall accuracy is 0.9083, indicating that it correctly predicts the class in 90.83% of instances. The model’s performance in all classes is measured through macro-average and weighted average accuracy, recall and F1-score values. These statistics offer a thorough assessment that accounts for the varied distribution of examples among classes.


       
Fig 5 shows the classifier’s ability to identify between many unhealthy classes, including early blight, late blight, leaf mold and a healthy class. Each unhealthy area under the curve (AUC) is displayed, with a greater AUC suggesting better performance in identifying the unhealthy from healthy plants. The AUC is low at 0.56 for early blight, showing the rate of false positives (FPR) or true positive rate (TPR) based on the classification level. In contrast, late blight had a higher AUC of 0.62, showing better distinction between the two.

Early detection of fungal infections allows the use of preventive treatments that involve antifungal treatments or crop rotation methods, resulting in reduced disease spread and production losses. Further research might focus on improving CNN designs and increasing the broad range of disease detection to include more crop species. Improvements in food security and environmentally friendly farming methods can be made by using deep learning skills and vast image datasets in the face of emerging challenges, such as pathogens, while also incorporating real-time tracking systems to improve precision agriculture.

We would like to thank LPPM Universitas Sam Ratulagi and all parties involved in the completion of this research.
 
Funding details
 
This research received no external funding.
 
Authors’ contributions
 
All authors contributed toward data analysis, drafting and revising the paper and agreed to be responsible for all the aspects of this work.
 
Data availability
 
The data analysed/generated in the present study will be made available from corresponding authors upon reasonable request.
 
Availability of data and materials
 
Not Applicable.
 
Use of Artificial Intelligence
 
Not applicable.
 
Declarations
 
Authors declare that all works are original and this manuscript has not been published in any other journal.
 
Disclaimers
 
The views and conclusions expressed in this article are solely those of the authors and do not necessarily represent the views of their affiliated institutions. The authors are responsible for the accuracy and completeness of the information provided but do not accept any liability for any direct or indirect losses resulting from the use of this content.

Authors declare that they have no conflict of interest.