Early detection and control of crop disease is essential for farmers, stakeholders, and precision agriculture researchers to reduce the production losses. Current farm practices rely on visual identification of plant diseases by farm staff with the backup of specialists using additional resources and tools, such as microscopes [1]. However, agricultural professionals cannot constantly be present in the field to perform thorough monitoring, and farmers lack the expertise required to conduct the detection procedure [2].

Multispectral, RGB, and hyperspectral sensors have been used for crop-disease detection [3]. Recently, crop-disease detection utilizing a variety of image sensors has shown encouraging results when combining data-driven approaches, such as machine learning (ML) and deep learning (DL) [4]. Tomatoes are a commercially significant vegetable crop on a global scale, and various pathogens (viral, bacterial, and fungal illnesses ^[5][6][7][8][5,6,7,8]) that affect tomatoes have been identified [9].

A number of researchers have focused on the use of classification models in disease diagnosis. The majority of the suggested classifiers are developed and validated, with an emphasis on extracting deep features from images in order to categorize the foliage disorders ^[10][11][10,11]. For instance, in an experiment, Trivedi et al. [12] classified nine different kinds of tomato leaf diseases using a convolution neural network and a dataset of 3000 tomato leaves. They attained an accuracy of 98.49% using pre-processed and segmented tomato leaf images.

Although the existing deep-learning models for tomato-leaf-disease recognition achieved high accuracy on selected leaf image datasets, their interpretability and explainability are not sufficiently investigated to engender trust in using such models in practice. The eXplainable Artificial Intelligence (XAI) and DL algorithms that produce human-readable explanations for AI judgments lay the groundwork for imaging-based artificial-intelligence applications [13] in various domains, such as health informatics [14], computer vision [15], and many more.

Given that DL-based learning may autonomously extract features from an image without the need for human feature engineering, it is vital to explain the model’s output in cases when the XAI can enhance it. A few studies have anticipated XAI with DL models for the prediction of different subtypes of tomato leaf diseases (TLD) to include explanatory results [16]. 

DL models have made significant advances in a variety of fields including, but not limited to, deep fakes ^[17][18][22,23], satellite image analysis ^[19][24], image classification ^[20][21][25,26], the optimization of artificial neural networks ^[22][23][27,28], the processing of natural language ^[24][25][29,30], fin-tech ^[26][31], intrusion detection ^[27][32], steganography ^[28][33], and biomedical image analysis ^[14][29][14,34]. CNNs have recently surfaced as one of the most commonly used techniques for plant disease identification ^[30][31][35,36].

By removing the constraints brought on by poor illumination and homogeneity in complicated environment scenarios, several works have concentrated their efforts on recognizing characteristics, while some authors have introduced real-time prediction ^[32][33][37,38]. For instance, research has been performed using DL models with the advancement of XAI techniques to develop a disease detection system with the major objective of pinpointing the disease and identifying the major areas of the plant and their parts that contribute to the classification ^[16][34][16,39].

The PlantVillage (PV) ^[35][40] dataset is a publicly available resource that contains images of various plant leaves with a range of disorders, including a tomato leaf disorder (TLD) ^[33][38]. This dataset has been used in multiple research works, including the following: Zhao et al. ^[36][41] achieved classification of TLDs using a multi-class feature-extraction approach. The residual block and the attention strategies were both integrated into the model, which was built on a deep CNN model. The model outperformed various deep-learning models with an overall accuracy of 99.24%.

Using the same image set, Bhujel et al. ^[37][42] examined the effectiveness of identifying various tomato diseases using a lightweight DL model. To enhance the performance of the model, a lightweight CNN method was combined with a number of attention strategies. The study explored the network architecture, performance, and computational complexity for the TLD dataset. The results showed improving classification accuracy upon building the compact and computationally efficient model with an accuracy of 99.69%.

TLD categorization was suggested by Ozbılge et al. ^[38][43] as an alternative to the well-known pre-trained knowledge-transferred ImageNet deep-network model and the compact deep-neural-network design with only six layers. The model’s performance on the PVdataset was tested using a number of statistical methods, and an accuracy of 99.70% was achieved. Antonio et. al. ^[39][44] suggested the use of a custom CNN-based architecture, which achieved a training accuracy of 99.99%, validation accuracy of 99.64%, precision of 99.00%, and a F1-score of 99.00% with the PV tomato dataset. With regard to the classification of the nine tomato illnesses, the recall metric had a value of 0.99.

The PlantVillage dataset was also utilized by Suryawati et al. ^[40][45] to train a model using Alexnet, GoogleNet, and VGGNet, which achieved test accuracies of 91.52%, 89.68%, and 95.25%, respectively. Transfer learning was used by Hong et al. ^[41][46] to reduce the quantity of training data needed, the amount of time typically required, and the cost of computation. Five deep-network topologies—Xception, Resnet50, MobileNet, ShuffleNet, and Densenet121—were employed to glean features from the 10 various tomato leaf disorders form the PV dataset. During the experiment, network architectures with various learning rates were contrasted. ShuffleNet had a recognition accuracy of 83.68%, whereas DenseNet and Xception had accuracies of 97.10% when the parameters were at their highest.

Vijay et al. ^[42][47] used CNN and K-nearest neighbor (KNN) models in their classification of tomato leaf disorders using the PlantVillage dataset, while LIME was used to provide explainability for the predictions made by each model. The CNN model performed better than the KNN model when used to detect leaf disease. The accuracy, precision, recall, and F1-score of the CNN model were 98.5%, 93%, 93% and 93%—all greater than those of the KNN model, which only managed to reach values of 83.6%, 90%, 84%, and 86%, respectively. Noyan et al. ^[43][48] claimed that the PlantVillage dataset is biased through the association of the background color to specific TLDs with an accuracy of around 40% for classification based on the use of background pixels only. However, Mzoughi et al. ^[44][49] demonstrated bias in the PV dataset, with image background colour associated with disease class. Additionally, they also showed improved identification outcomes, particularly in the setting of pictures with complicated backgrounds.

Kaur et al. ^[45][50] used an EfficientNetB7 model to examine leaf diseases of grape plants from the PlantVillage dataset. For the purpose of extracting the most important characteristics, the fully connected layer was created. The variance approach was then used to exclude extraneous features from the feature extractor vector. The logistic regression approach was then used to minimize the characteristics that had achieved a classification precision of 98.7%.