LeafDNet: Transforming Leaf Disease Diagnosis Through Deep Transfer Learning

ABSTRACT

The health and productivity of plants, particularly those in agricultural and horticultural industries, are significantly affected by timely and accurate disease detection. Traditional manual inspection methods are labor‐intensive, subjective, and often inaccurate, failing to meet the precision required by modern agricultural practices. This research introduces an innovative deep transfer learning method utilizing an advanced version of the Xception architecture, specifically designed for identifying plant diseases in roses, mangoes, and tomatoes. The proposed model introduces additional convolutional layers following the base Xception architecture, combined with multiple trainable dense layers, incorporating advanced regularization and dropout techniques to optimize feature extraction and classification. This architectural enhancement enables the model to capture complex, subtle patterns within plant leaf images, contributing to more robust disease identification. A comprehensive dataset comprising 5491 images across four distinct disease categories was employed for the training, validation, and testing of the model. The experimental results showcased outstanding performance, achieving 98% accuracy, 99% precision, 98% recall, and a 98% F1‐score. The model outperformed traditional techniques as well as other deep learning‐based methods. These results emphasize the potential of this advanced deep learning framework as a scalable, efficient, and highly accurate solution for early plant disease detection, providing substantial benefits for plant health management and supporting sustainable agricultural practices.

Abbreviations

AI

Artificial Intelligence

AUC

area under curve

BN

batch normalization

CNN

convolutional neural network

DL

deep learning

GPU

graphics processing unit

Grad‐CAM

gradient‐weighted class activation mapping

ML

machine learning

R‐CNN

region‐based convolutional neural network

ReLU

rectified linear unit

ROC

receiver operating characteristic

SVM

support vector machine

XAI

explainable AI

1. Introduction

Plants cultivated for ornamental, agricultural, or horticultural purposes are highly vulnerable to various diseases that adversely impact their health, productivity, and economic value. These diseases, which include bacterial, fungal, viral infections, and pest infestations, pose significant threats to global agriculture by diminishing crop yields, degrading quality, and escalating management costs. For instance, bacterial diseases such as bacterial spots in tomatoes and bacterial canker in mangoes result in considerable crop losses. Likewise, fungal diseases like anthracnose and powdery mildew affect mangoes and roses (Saleem, Potgieter, and Arif 2021, Albattah et al. 2022, Vishnoi, Kumar, and Brajesh 2021, Vallejo‐Pérez et al. 2021, Sharma, Mittal, and Gupta 2024b). Viral infections, including the tomato yellow leaf curl virus, further complicate these issues by inhibiting plant growth and reducing fruit quality (Vallejo‐Pérez et al. 2021).

The economic ramifications of these diseases extend beyond physical damage, leading to substantial financial losses for farmers due to decreased yields and the expenses associated with disease management. In severe cases, crop loss can be total, disrupting the overall food supply chain. Traditional disease detection methods, such as visual inspections and chemical testing, are labor‐intensive, prone to human error, and frequently delay the diagnosis and treatment of plant diseases (Sultana et al. 2023). A highly efficient convolutional neural network (CNN) model has been proposed for detecting diseases in potato leaves, providing a streamlined method for automating disease identification and minimizing human error, thereby enhancing precision agriculture (Lee et al. 2021). With the increasing global food demand, the need for scalable and efficient disease detection systems that facilitate timely interventions has become critical (Singh and Yogi 2023, Saleem, Potgieter, and Arif 2021).

This research presents a robust framework for diagnosing plant diseases using a deep neural network integrated with explainable AI (XAI) techniques, enabling transparent and interpretable detection in agricultural environments (Natarajan, Chakrabarti, and Margala 2024; Varur et al. 2023; Sharma, Mittal, and Gupta 2024b). These models enhance detection accuracy by identifying intricate visual patterns and adapting to evolving environmental conditions and disease manifestations, significantly outperforming conventional methods (Vishnoi, Kumar, and Brajesh 2021; Albattah et al. 2022). Deep learning (DL) approaches enable real‐time disease detection and have demonstrated their effectiveness in practical agricultural scenarios (Natarajan, Chakrabarti, and Margala 2024; Singh et al. 2020).

Despite advancements, many studies have concentrated on limited datasets or a restricted number of diseases. Our research aims to overcome these limitations by presenting a comprehensive approach that utilizes a multispecies dataset to identify bacterial, fungal, and pest‐related diseases across various crops.

Key Contributions of this study are as follows:

1. We propose a novel Xception‐based model utilizing transfer learning, fine‐tuned for classifying multiple leaf diseases across diverse plant species, including mangoes, roses, and tomatoes. This model integrates additional convolutional layers to enhance feature extraction, along with dense layers that include regularization and dropout techniques to boost accuracy and robustness.

2. A diverse collection of leaf images from multiple plant species was compiled. To increase variability, sophisticated data augmentation techniques, including rescaling, rotation, and flipping, were employed.

3. The proposed approach demonstrated exceptional performance, achieving 98% accuracy, 99% precision, 98% recall, and a 98% F1‐score, underscoring its superiority over traditional methods.

4. The proposed solution is scalable and suitable for implementation in diverse agricultural settings, enhancing early disease detection and minimizing yield losses.

5. Layer‐CAM and saliency maps were employed to visualize the model's decision‐making process, enhancing interpretability and fostering trust in its application for real‐world agricultural scenarios.

1.1. Background Study

Plant diseases pose a significant challenge to agriculture, adversely affecting both crop yield and quality. Essential crops like mango, tomato, and rose are especially susceptible to various pathogens, including bacterial, fungal, and viral infections, as well as infestations by pests. Traditional detection methods, such as visual inspections and chemical tests, are typically labor‐intensive, time‐consuming, and susceptible to errors. They highlight the need for automated solutions that can deliver fast and accurate disease diagnosis.

Recent advancements in DL, particularly CNNs, have revolutionized plant disease detection and classification. A cutting‐edge DL method leveraging CNNs has been developed to automatically extract features from plant images, demonstrating exceptional precision and robustness across various plant diseases. This method has demonstrated significant improvements over traditional techniques, providing an efficient solution for modern agricultural practices (Saleem, Potgieter, and Arif 2021). The approach's remarkable accuracy and generalization capabilities suggest its potential for widespread adoption in agricultural sectors (Albattah et al. 2022).

Additionally, Liu and Wang (2021) provide a review of DL applications in plant disease and pest detection, categorizing research into classification, detection, and segmentation networks. The study also identifies key challenges and suggests future research directions to drive progress in this field.

1.2. DL Approaches

CNNs have become the leading approach for image‐based plant disease detection, effectively extracting critical features from plant images to ensure accurate classification. Prominent architectures like ResNet, DenseNet, and EfficientNet have achieved accuracy rates exceeding 95% for diseases such as bacterial spot and powdery mildew (Singh et al. 2020; Sharma, Mittal, and Gupta 2024b). A stepwise DL model has been developed for plant disease detection, integrating both crop classification and disease identification, with a reported accuracy of 97.09%. Incorporating nonmodel crops into the training dataset has further improved accuracy, highlighting the model's potential in smart farming applications (Jung et al. 2023). Moreover, Rizwan et al. (2024) introduced a computationally efficient CNN model specifically designed to support small‐scale farmers. The model attained an average accuracy of 96.86% while functioning with limited computational resources, proving to be a practical solution for real‐world agricultural applications. Despite their high accuracy, the computational demands of these models often restrict their practical use in agricultural settings.

1.3. IoT and Edge Computing Integration

The combination of Internet of Things (IoT) sensors with edge computing devices has greatly enhanced real‐time, on‐site disease detection. Lightweight models, such as MobileNetV2, deployed on low‐power de‐ vices like Raspberry Pi, facilitate continuous monitoring and detection of diseases such as bacterial canker in mangoes and black spot in roses.

In the context of precision agriculture, accurate and continuous monitoring of plant health is critical for enhancing crop quality and food security.

Research by Li et al. (2024) explores the potential of wearable sensors for real‐time monitoring of plant health, providing insights into physiological biomarkers. The study provides a detailed overview of wearable plant sensors, covering their design, sensing mechanisms, and applications, thereby guiding the progress of plant health monitoring in precision agriculture. Furthermore, utilizing attention mechanisms has demonstrated effective disease detection, achieving accuracy above 90%, thus reducing labor costs associated with traditional inspection methods Sultana et al. (2023).

Research on standalone Edge‐AI solutions for tomato leaf disease detection has demonstrated the efficiency of lightweight DL architectures, such as GoogleNet and MobileNetV2, which have achieved impressive accuracy rates of up to 98.25%. These studies highlight the potential of compact and resource‐efficient models for real‐time agricultural applications. Additionally, this research underscores the effectiveness of edge devices like NVIDIA Jetson and Raspberry Pi in facilitating real‐time plant disease detection, showcasing their practical value in modern agricultural settings (Majeed, Ojo, and Zahid 2024). Expanding on this, Nayagam et al. (2023) proposed a parallel and distributed simulation framework (PDSF) designed for pest management and crop monitoring. By utilizing multiple GPUs, the framework achieved remarkable performance metrics of 98.65%, significantly improving the efficiency of pest control and crop surveillance operations, and demonstrating its applicability for large‐scale agricultural management.

1.4. Spectral and Hyperspectral Imaging

Beyond CNN‐based models, spectral imaging techniques, including hyperspectral and near‐infrared (NIR) imaging, have proven effective for early disease detection, even before visible symptoms manifest. These methods are adept at identifying subtle physiological changes in plants, which can improve intervention strategies. Studies employing hyperspectral imaging to detect powdery mildew in roses and mangoes have reported detection accuracies exceeding 90% (Vallejo‐Pérez et al. 2021). Furthermore, early identification of viral infections through multispectral imaging–spectral machine learning has shown promise in enhancing diagnostic accuracy (Peng et al. 2022). A review of next‐generation methods for early disease detection discusses a range of traditional and innovative techniques, including DNA, RNA, and protein‐ based methods, alongside portable PCR assays and lateral flow systems to improve pathogen detection efficiency and accuracy in field settings (Trippa, Scalenghe, and Basso 2023).

1.5. Multidisease Classification and Transfer Learning

In practical agricultural contexts, it is common for plants to suffer from multiple concurrent diseases, necessitating multidisease classification models capable of identifying various diseases in a single analysis. Recent advancements in deep transfer learning have yielded promising results in detecting tomato leaf diseases, enabling models to classify diseases such as bacterial spots and late blight with high precision.

Nawaz et al. (2022) introduced a robust methodology for localizing and classifying tomato leaf diseases using a ResNet‐34‐based Faster‐RCNN model, achieving an impressive mean average precision (mAP) of 0.981 and an accuracy of 99.97% under diverse image conditions. Transfer learning, which adapts pre‐trained models to specific datasets, has been extensively employed to improve classification accuracy. Kabala et al. (2023) explored the use of federated learning for crop disease detection, enabling collaborative model training without compromising sensitive data.

Additionally, Sharma et al. (2024a) proposed a stacked DL framework for multiclass classification of plant diseases, leveraging an ensemble of advanced DL algorithms combined with a multi‐layer perception as a meta‐classifier, achieving a peak accuracy of 98.13%. Khan et al. (2024) developed seven Bayesian‐optimized hybrid DL models for automated tomato leaf disease classification, integrating CNNs for feature extraction with traditional machine learning techniques and stacking for classification, ensuring both time efficiency and generalizability. These models have successfully detected diseases such as anthracnose in mangoes and downy mildew in sunflowers (Varur et al. 2023; Fuentes et al. 2021). Tiwari, Joshi, and Dutta (2022) conducted an extensive study utilizing a diverse dataset comprising 12 crops and 22 categories, achieving an average cross‐validation accuracy of 98.68% and a test accuracy of 97.69% on unseen images. This highlights the model's capability to handle both intra‐class and inter‐class variations effectively.

1.6. Responsible AI (RAI) and XAI in Plant Disease Detection

RAI emphasizes the ethical, transparent, and beneficial development of AI systems. In plant disease detection, it is imperative that models not only demonstrate accuracy but also maintain fairness, explainability, and accountability when deployed in real‐world agricultural contexts. XAI techniques, such as LayerCAM and Saliency Maps, enhance the interpretability of DL models by offering visual explanations of the reasoning behind their predictions. Recent literature provides an extensive overview of XAI, discussing its concepts, taxonomies, opportunities, and challenges. The emphasis on explainability is critical for the successful deployment of AI models, ensuring reliable explanations to mitigate decision‐making errors. Moreover, the limitations of XAI are explored alongside suggestions for future research directions (Yang et al. 2023).

Integrating XAI techniques into AI models for plant disease detection fosters transparency, allowing stakeholders to identify potential biases and make informed decisions, thereby preventing erroneous or harmful outcomes. This is especially vital for identifying diseases requiring specific treatment protocols, as misdiagnosis can result in unnecessary pesticide use or crop loss (Tjoa and Guan 2021). Ensuring fairness in model performance across different crop types and disease categories is essential to avoid biases arising from imbalanced datasets.

Recent studies on RAI in agriculture emphasize the importance of tackling concerns related to data privacy, bias reduction, and the socio‐environmental effects of AI implementation in resource‐limited regions. The study by Mana et al. (2024) explores the applications and implications of AI in sustainable agriculture, highlighting its potential to tackle issues such as climate change and food insecurity. It discusses how AI can enhance precision agriculture, optimize resource use, and integrate renewable energy solutions to support global sustainability goals. Furthermore, Adve et al. (2024) discuss the AIFARMS initiative, a collaborative effort involving over 40 AI and agriculture researchers aimed at developing AI technologies for sustainable and resilient agriculture, with a focus on areas such as autonomous farming, cover crop planting, machine learning for remote sensing, yield loss estimation, and livestock management. AIFARMS also emphasizes the creation of high‐quality, open data sets and the cultivation of a skilled, diverse workforce in digital agriculture.

1.7. Future Directions and Overall Findings

The literature suggests that the integration of DL, IoT, and spectral imaging is transforming plant disease detection. Models such as DenseNet and MobileNet have shown impressive accuracy; however, several challenges remain, including the scarcity of large datasets, the need for computational efficiency, and issues surrounding model generalization. The adoption of RAI and XAI methods is becoming increasingly essential to ensure transparency, fairness, and ethical use of these models. XAI techniques are pivotal in making AI systems understandable to farmers and stakeholders, especially during critical decision‐making processes involving disease diagnosis and treatment recommendations.

Future research should prioritize the integration of XAI methods to enhance trustworthiness and accountability within AI systems. Furthermore, developing scalable and robust models that can be effectively deployed in resource‐constrained environments is essential. Addressing the challenge of creating more diverse datasets that encompass various plant species and diseases is also imperative. An overall summary is discussed in Table 1

TABLE 1.

Overview of key findings from literature on plant disease detection.

Key findings	Implications for future research
High accuracy of deep learning models	Focus on integrating diverse datasets
Importance of real‐time monitoring	Explore applications of AI in IoT settings
Need for explainability in AI models	Develop frameworks for model interpretability
Field testing of models	Conduct trials in diverse agricultural contexts
Emergence of responsible AI practices	Address ethical and sustain‐ability considerations

The paper is structured into four main sections. Section 1, the Introduction, offers an overview and reviews pertinent literature. Section 2, Materials and Methods, outlines the materials used in the study and provides a detailed explanation of the methods, with a focus on the Xception model and its implementation. Section 3, Results, presents the experimental findings and their evaluation. Finally, Section 4, Discussion, examines the implications of the results, summarizes the study, and suggests potential avenues for future research.

2. Materials and Methods

2.1. Dataset Overview

A robust and diverse dataset is pivotal for evaluating the efficacy of recognition algorithms throughout the research process. In this study, we utilized five distinct datasets to comprehensively assess the model's performance across various plant diseases. Table 2 provides an overview of the datasets used in the study, including the FlowerNet dataset, ISPGULDataset8C, Chinese Rose Leaf Disease Dataset, Mango Leaf Disease Dataset, and Tomato Leaf Disease Dataset.

TABLE 2.

Overview of collected datasets.

Name	Classes	Description	Total images
FlowerNet (Rajbongshi et al. 2022)	Black spot, downy mildew, and new leaf	A dataset comprising images of both healthy and diseased rose leaves, emphasizing common diseases like black spot and downy mildew. This dataset includes both original images (917) and augmented datasets (4342) to enhance model training. Images are sized at 512 × 512 pixels and sourced from botanical settings in Dhaka, Bangladesh.	5259
ISPGULDataset8C (Duman 2023)	Multiple diseases	Contains 567 images of damask rose plants in natural conditions, showcasing a variety of diseases and pest infestations, thus providing a real‐world context for model evaluation.	567
Chinese rose leaf disease dataset (Zhang and QI 2023)	Multiple diseases	The dataset comprises images of both healthy and diseased leaves, which have been randomly augmented and partitioned for research purposes. This approach enables a thorough evaluation of classification algorithms, ensuring a comprehensive assessment of their performance.	3326
Mango leaf disease dataset (Shakib, Mustofa, and Ahad 2024)	Multiple diseases	Comprises 4000 images of mango leaves, focusing on seven diseases along with healthy samples, collected from various orchards across Bangladesh, thereby ensuring geographical diversity.	4000
Tomato leaf disease dataset (Dalal, Lal, and Patel 2024)	Multiple Diseases	The dataset comprises 5000 images of tomato leaves in different health conditions, collected under natural environments, enabling a realistic assessment of the model's ability to detect diseases.	5000

The datasets used in this study were meticulously curated to encompass a diverse range of diseases across three distinct and representative crop types: fruits (mango), vegetables (tomato), and flowers (rose). This deliberate selection strikes a balance between covering key agricultural and horticultural domains while considering the need for additional plant species. Geographic conditions, while valuable for enhancing model generalizability, introduce substantial variability due to differences in disease manifestations influenced by climate, soil, and regional practices. Including such variability could complicate the model's adaptation to real‐time, cutting‐edge technologies. Moreover, addressing geographic diversity might require supplementary data collection beyond images, creating challenges in achieving optimal performance with fewer parameters. This study carefully controlled variability to ensure a thorough and robust evaluation of the model's architecture, emphasizing its capacity to generalize effectively across a wide range of major crop types and diverse environmental conditions. To achieve this, the dataset was systematically divided into three subsets: 3880 images (70%) were allocated for training, 814 images (15%) for validation, and 817 images (15%) for testing. This partitioning maintained a uniform class distribution across all subsets, ensuring the reliability of the performance assessment. A comprehensive summary of the dataset composition can be found in Table 3. The dataset includes the following categories:

• Bacterial Diseases: It includes Bacterial Spot (Tomato) and Bacterial Canker (Mango), both of which can severely impact plant health and yield.

• Fungal Diseases: It includes prevalent diseases such as Anthracnose (Mango), Powdery Mildew (Rose, Mango), Downy Mildew (Rose), Sooty Mould (Mango), and Leaf Mold (Tomato), all of which can cause considerable crop losses if not addressed in a timely manner.

• Pests and Infestations: It covers damage caused by pests such as Gall Midge (Mango), Cutting Weevil (Mango), Macrosiphum rosae (Rose), and Spider Mites (Tomato), emphasizing the importance of pest management in agricultural practices.

• Healthy Leaves: It represents healthy samples of Rose, Tomato, and Mango leaves, serving as a baseline for comparison against diseased samples.

TABLE 3.

Dataset composition by class and partition.

Class	Training set (70%)	Validation set (15%)	Testing set (15%)
Bacterial diseases	828 images	177 images	177 images
Fungal diseases	426 images	91 images	91 images
Pests and infestations	923 images	182 images	184 images
Healthy leaves	1703 images	364 images	365 images

Note: The bold rows represent the accuracy of the proposed model in this research, which supports the novelty of the study.

2.2. Dataset Compilation and Enhancement

The core of this study lies in a meticulously curated dataset aimed at classifying diseases affecting rose foliage. The dataset includes images from multiple sources and environments to represent variations in disease appearance and environmental conditions. The dataset, consisting of 5511 images, was deliberately curated to encompass a wide variety of disease types and plant categories, thereby improving the model's applicability and robustness across diverse agricultural contexts. The images were meticulously divided into separate subsets for training, validation, and testing to enable a comprehensive assessment of the model's performance.

2.3. Dataset Segregation

The dataset was divided into training, validation, and testing sets as follows:

• Training Set: 70% (3880 images) used to train the model, allowing it to learn from various examples.

• Validation Set: 15% (814 images) employed to fine‐tune model parameters and prevent overfitting during training.

• Testing Set: 15% (817 images)—This subset was utilized to evaluate the final model's performance and its capability to generalize to unseen data.

2.4. Data Generation

Using the flow‐from‐directory method provided by ImageDataGenerator, we generated augmented image batches for training and validation. The training generator was configured to shuffle images, enhancing model robustness and improving generalization. The validation generator processed images sequentially to maintain evaluation consistency. The test set was normalized but not augmented to preserve evaluation integrity, ensuring that performance metrics are accurately reflective of the model's capability to classify real‐world images.

2.5. Class Composition

The dataset encompasses four unique classes, each representing a distinct disease or condition affecting rose leaves. The diversity within the training set was verified, ensuring all categories were adequately represented, as detailed in Table 4.

TABLE 4.

Class distribution across the dataset.

Class	Number of images
Bacterial diseases	1182 images
Fungal diseases	608 images
Pests and infestations	1289 images
Healthy leaves	2432 images

2.5.1. Image Preprocessing and Augmentation

In deep learning, particularly in computer vision, image preprocessing and augmentation are vital steps that significantly contribute to improving model performance. These techniques enhance both the quality and variety of the dataset while also aiding in the prevention of overfitting, thereby ensuring that the model generalizes well to new, unseen data. Below, we explain each technique employed in our methodology, its general utility, and the specific parameters applied to our dataset.

• Rescaling: This technique plays a vital role in normalizing pixel values in images, ensuring consistency and stability during the model's learning process. By scaling pixel intensity values to a standardized range, typically between [0, 1], the model can process inputs more efficiently and achieve better convergence. In this study, normalization was performed by dividing each pixel value by 255, the maximum intensity value for an 8‐bit color channel, thereby standardizing the input data for optimal performance.

• Rotation: To account for the variations in orientation that may occur naturally, the images were randomly rotated. This introduces rotational invariance to the model, allowing it to recognize patterns irrespective of their orientation. To replicate this effect, the images were randomly rotated within a 10‐degree range.

• Translation: Images were randomly shifted horizontally and vertically to simulate the effect of objects in different positions within the frame. This teaches the model to focus on the features rather than their position in the image. We applied shifts of up to 10% in image width and height.

• Shearing: Shear transformation reduces the shape of the image, simulating a change in perspective. This is particularly useful for training models that recognize objects at various poses and angles. Our images underwent shear transformations with intensities of up to 0.2 radians.

• Zooming: Random zooming either crops or expands images, which can mimic the effect of objects at different distances from the camera. This variation helps the model maintain accuracy regardless of the scale. The zoom range was set to 30%, allowing for significant scale variation.

• Flipping: Mirroring images horizontally can effectively double the size of a dataset and introduce additional variability. This is particularly useful for objects that are symmetrical or where the orientation is not a defining characteristic. In our dataset, images randomly flipped along the horizontal axis.

• Filling: After applying transformations, such as rotation or shifting, new pixels are introduced into the image. The filling strategy determined the population of these pixels. We used the “nearest” strategy, which populates the new pixels with the nearest existing pixel values, maintaining the local structure of the image.

Each technique was meticulously designed to capture the natural variability observed in real‐world scenarios, enriching the dataset and improving the robustness of the classification model. All images were resized to a consistent dimension of 256 × 256 pixels, a resolution selected for the reasons explained in the Results section and were normalized to facilitate efficient training.

2.6. Proposed Model

The proposed model architecture is built upon the Xception pretrained model, which is renowned for its depth wise separable convolutions and efficient feature extraction capabilities (Chollet 2017). The Xception model, pretrained on the ImageNet dataset (Russakovsky et al. 2015), served as the backbone of our architecture, providing robust feature maps that are crucial for subsequent layers. The proposed model architecture is illustrated in Figure 1.

FIGURE 1.

The proposed method for classifying rose leaf diseases at different stages includes preprocessing, a pretrained Xception model, and several custom layers.

2.6.1. Base Model:Xception

The Xception architecture represents a significant advancement in CNN design, utilizing depthwise separable convolutions to reduce computational complexity while maintaining strong representational power. As described by Chollet (2017), Xception is built upon the principle of mapping spatial correlations separately from channel‐wise correlations, enhancing the network's ability to capture complex patterns. Xception's architecture can be mathematically described as follows:

$$y=H\left(x\right)=f\text{depthwise}\left(f\text{pointwise}\left(x\right)\right)$$ (1)

where f _depthwise represents the depthwise convolution, which applies a single filter per input channel, and f _pointwise refers to the pointwise convolution, which combines these outputs using 1 × 1 convolutions. This structure leads to improved computational efficiency while preserving model accuracy.

The Xception base model, excluding its top layer, was used with an input shape of (256 × 256 × 3). The model was loaded with pretrained ImageNet weights, ensuring that the initial feature extraction leverages learned representations from a large and diverse dataset. The mathematical representation of the output feature maps from the Xception model is given by the following:

$${\mathbf{F}}_{\text{base}}=\text{Xception}\left(\mathbf{X}\right)$$ (2)

where X denotes the input image tensor.

2.6.2. Convolutional Layers

A convolutional layer is a fundamental part of a CNN, mainly aimed at detecting patterns like edges, colors, gradients, and other visual elements in images. It uses learnable filters, or kernels, that slide over the input image or feature map, producing a transformed feature map through a process known as convolution. The convolution operation in a convolutional layer is generally represented mathematically as:

$${\mathbf{F}}_{\text{out}}=\text{Activation}\left(\text{BatchNorm}\left(\text{Conv}2{\mathrm{D}}_{k,f\times f}\left({\mathbf{F}}_{\text{in}}\right)\right)\right)$$ (3)

where:

• F _in is the input feature map that contains the initial representation of the image or features from the previous layer.

• Conv2D _k,f × f represents the convolution operation with k filters of size f × f, which slide over F _in to produce output feature maps.

• BatchNorm normalizes the output of Conv2D to stabilize learning by reducing internal covariate shift.

• Activation refers to the activation function applied to the normalized feature map, with the Rectified Linear Unit (ReLU) being one of the most commonly used functions, which introduces non‐linearity.

• F _out is the transformed output feature map that represents higher‐level features learned by the convolutional operation.

  For the first convolutional layer after Xception:

  $${\mathbf{F}}_1=\text{BatchNorm}\left(\text{ReLU}\left(\text{Conv}2{\mathrm{D}}_{\mathrm{256,3}\times 3}\left({\mathbf{F}}_{\text{base}}\right)\right)\right)$$ (4)

Here, F _base serves as F _in, the output from Xception. Conv2D_256,3×3 applies 256 filters of size 3 × 3, followed by ReLU activation and batch normalization, enhancing feature extraction and robustness.

For the second convolutional layer:

$${\mathbf{F}}_2=\text{BatchNorm}\left(\text{ReLU}\left(\text{Conv}2{\mathrm{D}}_{\mathrm{128,3}\times 3}\left({\mathbf{F}}_1\right)\right)\right)$$ (5)

F ₁ from the first layer acts as F _in, and Conv2D_{128,3 × 3} applies 128 filters of size 3 × 3, further refining feature maps for subsequent layers.

2.6.3. Global Average Pooling (GAP) Layer

GAP reduces the spatial dimensions of feature maps by averaging their values, focusing on key features and improving the model's simplicity and generalization.

$$\mathrm{GAP}\left(x\right)=\frac{1}{N}{\sum }_{i=1}^N{x}_i$$ (6)

where

• x _i are elements of x, representing features from the last convolutional layer.

• N represents the total number of elements in xxx, enabling the feature map to be averaged and spatial information to be condensed.

$${\mathbf{F}}_{\mathrm{GAP}}=\text{GlobalAveragePooling}2\mathrm{D}\left({\mathbf{F}}_2\right)$$ (7)

Here, F2 acts as the input (x) in the GAP layer, reducing the spatial dimensions of the feature maps while preserving the critical information needed for classification.

2.6.4. Fully Connected (FC) Layers

The proposed neural network architecture consisted of a series of three FC layers, each with a gradually decreasing number of neurons. These layers are essential for integrating high‐level features extracted by the earlier convolutional layers into a format suitable for classification.

2.6.4.1. Dense Layer Configuration

Each FC layer is followed by batch normalization and a dropout technique, which collaborate to minimize the risk of overfitting and improve the model's ability to generalize.

The configuration for each dense layer is as follows:

$${\mathbf{D}}_i=\text{Dropout}\left({\alpha }_i\right)\left(\text{BatchNorm}\left(\text{ReLU}\left({\text{Dense}}_{u}i\left({\mathbf{D}}_{i–1}\right)\right)\right)\right)$$ (8)

where D _i denotes the output of the i‐th dense layer, α _i is the dropout rate, u _i represents the number of units in the i‐th layer, and D _i–1 is the input received from the previous layer or the feature map.

2.6.4.2. Activation Function:ReLU

The ReLU (Nair and Hinton 2010) activation function is employed within the FC layers owing to its efficacy in maintaining the non‐linearity of the model:

$$\text{ReLU}\left(x\right)=\max \left(0,x\right)$$ (9)

This function facilitates a more expeditious training phase and endows the network with the capability of assimilating complex patterns by setting negative inputs to zero.

2.6.4.3. Regularization Techniques

L2 regularization and dropout are employed to further improve the model's generalization capability:

$$\mathrm{L}2\left(\mathbf{w}\right)=\mathrm{\lambda }{\sum }_{i=1}^n{w}_i^2$$ (10)

$$\text{Dropout}\left(x\right)=x\odot M$$ (11)

In the context of L2 regularization, λ is the regularization coefficient, and w _i is the weight of the network. This regularization form encourages the model to prefer smaller weights, thereby averting overfitting.

For dropout, x symbolizes the input to the layer, and M is a binary mask that is independently sampled for each neuron during the training iterations. This technique hinders the co‐adaptation of neurons, thus enhancing the robustness of the model.

2.6.4.4. Normalization Technique:Batch Normalization

Batch normalization, as introduced by Ioffe and Szegedy (2015), is applied after each activation function to normalize layer activations. This process helps accelerate convergence and reduces the internal covariate shift:

$$\text{BatchNorm}\left(x\right)=\mathrm{\gamma }\left(\frac{x-\mathrm{\mu }}{\sqrt{{\mathrm{\sigma }}^2+\mathrm{ϵ}}}\right)+\beta$$ (12)

Here, γ and β are trainable parameters, while μ and σ² represent the batch's mean and variance, respectively. ϵ is a small constant added to avoid division by zero.

2.6.4.5. Dense Layer Specifics

The first dense layer comprised 128 neurons, the second 64 neurons, and the third 16 neurons, each applying ReLU activation and L2 regularization. The dropout rate was set at 0.1 for all layers. The equations for the dense layers are as follows:

$${\mathbf{D}}_1=\text{Dropout}\left(0.1\right)\left(\text{BatchNorm}\left(\text{ReLU}\left({\text{Dense}}_{128}\left({\mathbf{F}}_{\mathrm{GAP}}\right)\right)\right)\right)$$ (13)

$${\mathbf{D}}_2=\text{Dropout}\left(0.1\right)\left(\text{BatchNorm}\left(\text{ReLU}\left({\text{Dense}}_{64}\left({\mathbf{D}}_1\right)\right)\right)\right)$$ (14)

$${\mathbf{D}}_3=\text{Dropout}\left(0.1\right)\left(\text{BatchNorm}\left(\text{ReLU}\left({\text{Dense}}_{16}\left({\mathbf{D}}_2\right)\right)\right)\right)$$ (15)

These dense layers are essential for enabling the network to identify complex patterns and features. L2 regularization plays a key role by penalizing large weights, promoting the learning of more generalized features.

2.6.5. Output Layer

In multiclass classification tasks, the softmax layer, placed as the final layer, plays a crucial role. It transforms the output logits into a probability distribution across the predicted output classes:

$$\text{Softmax}\left(x_i\right)=\frac{e^{x_i}}{\sum _j{e}^{x_j}}$$ (16)

where

• x _i are the logits (raw outputs) from the previous layer.

• $\sum ^{e^{x_i}}$ computes the exponential of x _i , ensuring all values are positive.

• _j e ^xj sums the exponentials across all classes, normalizing the distribution.

  $${\mathbf{Y}}_{\text{pred}}=\text{Softmax}\left({\text{Dense}}_4\left({\mathbf{D}}_3\right)\right)$$ (17)

Here, D ₃ provides logits as x _i , and dense₄ transforms them into 4 output classes, applying the softmax function to generate probabilities.

2.6.6. Model Summary

The architecture integrates a pretrained Xception model with additional convolutional layers, dense layers, and regularization techniques. This design leverages pretrained knowledge while allowing customization for specific classification tasks. The model's architecture is optimized for effective feature extraction and accurate classification, incorporating regularization and dropout to prevent overfitting and enhance generalization. The model consists of the following components:

• Xception Base: Pretrained on the ImageNet dataset, offering a robust set of initial features.

• Convolutional Layers: Two Conv2D layers with 256 and 128 filters, using ReLU activation and 'same' padding, followed by Batch Normalization.

• Global Average Pooling Layer: Aggregates feature maps to reduce spatial dimensions

• Dense Layers: Three customized dense layers with 128, 64, and 16 units, each employing ReLU activation and L2 regularization with a coefficient of 0.0001.

• Batch Normalization and Dropout: Applied following each dense layer to stabilize training and mitigate overfitting, utilizing a dropout rate of 0.1.

• Output Layer: A dense layer with a softmax activation function to classify inputs into four distinct categories.

2.6.7. Training Approach

We used the Adam optimizer with a learning rate of 1 × 10^–4 and categorical cross‐entropy as the loss function to promote efficient model convergence. The dataset was meticulously curated and organized with a batch size of 16 to maintain stability during training.

Overfitting was mitigated by using early stopping, which terminated training if the model's performance did not improve for three consecutive epochs. Furthermore, model checkpoints were utilized to save the best‐performing model during training.

The best‐performing model was saved based on validation loss. The training parameters, as detailed in Table 5, include a range of state‐of‐the‐art models carefully tuned for optimal performance on the NVIDIA Tesla T4 GPU.

TABLE 5.

Training specifications.

Parameter	Value
Optimizer	Adam
learning rate	$1\times {10}^{-4}$
Loss function	Categorical cross‐entropy
Batch size	16
Epoch	100
Early stopping patience	3
GPU	NVIDIA Tesla T4
Libraries	Keras, TensorFlow

A focused training strategy of 100 epochs was applied, prioritizing efficiency while incorporating early stopping. The training process was accelerated by using multiple workers and leveraging the computational power of the NVIDIA Tesla T4 GPU on Google Colab, ensuring quick and effective model optimization.

The Adam optimizer, with a learning rate of 1 × 10^–4, determined through extensive experimentation, along with categorical cross‐entropy loss, formed the core of our training strategy. The dataset, consisting of both original and augmented images, was carefully curated and organized with a batch size of 16 to maintain stability during training.

2.7. XAI

Conventional methods for disease detection, such as visual inspections and chemical testing, are time‐consuming, susceptible to human error, and often result in delayed diagnosis and treatment of plant diseases.

However, recent developments in deep learning, particularly CNNs, have revolutionized plant disease detection by automating the process of identification, offering greater accuracy and reliability. In this study, we utilized a modified Xception model and incorporated XAI techniques such as LayerCAM and Saliency Maps to achieve high accuracy while ensuring interpretability. The model was pretrained and fine‐tuned on a varied dataset containing images of roses, mangoes, and tomato leaves, both diseased and healthy. These were classified into four categories:bacterial diseases, fungal diseases, pests and infestations, and healthy leaves. Each class was represented equally to ensure balanced learning and mitigate biases in disease classification.

2.7.1. LayerCAM

LayerCAM, which stands for Layer‐wise Contextual Attribution Maps, is an advanced technique that builds upon Grad‐CAM to provide visual explanations the decisions made by CNNs (Jiang, Zhang, and Wang 2021). This method calculates importance weights for each feature map within a specific layer, which are then utilized to create a heatmap overlay on the input image. The equations for the LayerCAM are as follows:

$$w_k^c=\sum _i\sum _j{\alpha }_{\mathit{ijk}}^c$$

$$\text{ReLU}\left(\frac{\partial A_{\mathit{ijk}}}{\partial Y_c}\right)$$

$${\mathrm{L}}_{\mathrm{ij}}^{\mathrm{c}}=\sum _{\mathrm{k}}{\mathrm{w}}_{\mathrm{k}}^{\mathrm{c}}·{\mathrm{A}}_{\mathrm{ijk}}$$

Here, w ^c represents the importance weight of the k‐th feature map for class c, A _ijk is the activation of the k‐th feature map at spatial location(i, j), and α ^c are the pixel‐wise coefficients for the gradients. The term ^∂Aijk denotes the gradient of the activations with respect to the class score ^Y c, and the ReLU function is used to ensure that only positive influences on the class score are taken into account. The resulting Lc values generate a heatmap that highlights the critical regions for class.

2.7.2. Saliency Maps

Saliency maps, introduced by Simonyan, Vedaldi, and Zisserman (2014), are a technique used to highlight the most significant pixels contributing to a model's prediction. This is accomplished by calculating the gradient of the output class score relative to the input image. The formula for generating a saliency map is as follows:

$$M_c\left(x\right)=\frac{\partial f_c\left(x\right)}{\partial x}$$

where $M_c\left(x\right)$ represents the saliency map for class c and $\frac{\partial f_c\left(x\right)}{\partial x}$ is the gradient of the class score $f_c\left(x\right)$ with respect to the input image $x$ is computed. Saliency maps offer an intuitive visualization, highlighting the areas of the image that are most crucial for the classification decision.

The implementation involved initializing the Xception‐based model and organizing the image datasets into their respective categories. Each image was preprocessed and fed into the model to obtain predictions. The LayerCAM and SaliencyMAP techniques were then applied to generate visual explanations of these predictions. The comparison between actual and predicted labels, combined with visual explanations, provided valuable insights into the model's performance, improving the transparency and accuracy of its predictions. Section 3 further expands on the detailed findings of this study.

3. Results

In this study, we present a customized Xception‐based classification approach aimed at identifying and classifying different stages of leaf diseases. Evaluation using several performance metrics, such as accuracy, precision, and recall, shows that the optimized model significantly outperforms the baseline Xception model, especially in terms of macro‐average values across these metrics. Despite this, there is room for improvement across various metrics, which we addressed by introducing specific modifications to the Xception architecture. These modifications effectively mitigated the identified limitations and enhanced overall performance. To rigorously confirm these performance improvements, tests were conducted on the outcomes of the testing dataset, later depicted as confusion matrices. The testing results confirmed that the observed differences between the proposed model and baseline models were statistically significant, further reinforcing the reliability of the claims. As illustrated in Figure 2, the confusion matrices provided clear insights into the model's prediction accuracy, with a strong focus on correctly predicted instances, highlighted by the intensity along the diagonal of the matrix.

FIGURE 2.

Confusion matrix from popular models we generated higher accuracy.

Additionally, Figures 3, 4, 5, 6, 7, 8, 9 offer a thorough validation of the model's performance, utilizing statistical metrics to assess its effectiveness in detecting leaf diseases.

FIGURE 3.

Comparative accuracy and loss metrics in training and validation.

FIGURE 4.

Comparative training and validation curves for precision and recall.

FIGURE 5.

This graphic illustrates the AUC score for training vs. validation curve.

FIGURE 6.

Loss of each CNN model.

FIGURE 7.

Comparison of precision and recall across different CNN models.

FIGURE 8.

The accuracy's of each model.

FIGURE 9.

The performance of each class for the proposed method is succinctly summarized in this bar chart.

Figure 3 highlights a significant improvement in accuracy during both the training and validation phases, supported by statistical validation to ensure consistency. The convergence of training and validation accuracy metrics, along with a steady upward trend, demonstrates the model's robust predictive capabilities. Furthermore, the validation loss consistently being lower and closer to the training loss indicates effective generalization, minimizing the risk of overfitting.

As depicted in Figure 4, the precision and recall metrics are consistently higher for the validation set compared to the training set. This further confirms the reduced likelihood of overfitting, leading to optimal model performance. The detailed statistical analysis demonstrates that the enhanced validation precision and recall highlight the model's ability to effectively balance true positive identification while minimizing false negatives, making it highly effective for multiclass classification tasks.

Figure 5 illustrates the receiver operating characteristic (ROC) curve alongside the area under the curve (AUC) score for the proposed model. The AUC value, which approaches 1, underscores the model's exceptional ability to distinguish between various leaf disease classes. This high AUC score further validates the model's effectiveness in accurately identifying and classifying different stages of plant diseases.

Figure 6 presents a comparative analysis of loss values across various CNN models, including the proposed model. The proposed model achieves the lowest loss value, indicating its superior predictive performance compared to the other models. Although Xception also demonstrates relatively low loss, models such as InceptionResNetV2 and DenseNet201 perform competitively with slightly higher loss values. On the other hand, models like NASNetMobile, MobileNetV2, NASNetLarge, ResNet152V2, and InceptionV3 display significantly higher loss values, reflecting weaker predictive capabilities. The minimal loss observed in the proposed model highlights its robust performance across metrics such as accuracy, precision, recall, and F1‐score.

Figure 7 provides a scatter plot comparing the performance of various CNN models using precision and recall metrics. The proposed model, marked by a red cross, clearly outperforms the others with precision and recall values of 0.99 and 0.98, respectively. Xception, represented by a light blue circle behind ResNet152V2's green square, also performs well with a precision of 0.96 and a recall of 0.95, showcasing its strength in recall‐oriented tasks. ResNet152V2, depicted as a green square, demonstrates moderate but consistent performance with precision and recall both at approximately 0.95.

InceptionV3, shown as an orange triangle, achieves balanced performance with precision and recall values near 0.94. Other models, including MobileNetV2, NASNetMobile, and NASNetLarge, exhibit slightly lower precision and recall, around 0.93, indicating higher misclassification rates. VGG19, represented by a gray diamond, records the lowest performance, with precision and recall both around 0.92. These results emphasize the superior architecture and optimization of the proposed model while providing valuable insights into the relative strengths and weaknesses of each CNN model for image classification tasks.

The performance assessment of different CNN models, including NASNetMobile, NASNetLarge, VGG19, MobileNetV2, ResNet152V2, InceptionV3, Xception, InceptionResNetV2, and DenseNet201, reveals notable differences in their accuracy, precision, recall, and F1‐scores. The proposed model outperformed all others, achieving an impressive accuracy of 0.98, followed by Xception at 0.96 and ResNet152V2 at 0.95. The NASNetMobile model recorded a lower accuracy of 0.94, while VGG19 was the lowest performer among the models at 0.92.

The proposed model achieved exceptional performance, with precision (0.99), recall (0.98), and F1‐score (0.98), highlighting its robust capabilities across key metrics. While DenseNet201 and InceptionV3 performed well with high scores around 0.95, they still fell short of the proposed model. Models like MobileNetV2, InceptionResNetV2, and NASNetMobile demonstrated relatively low yet competitive scores.

This performance underscores the significant advantages of the proposed model's architecture and optimization strategies, which enable it to handle diverse and complex datasets more effectively than other models, as illustrated in Table 6 and Figure 8.

TABLE 6.

Performance evaluation of applied methods with different accuracy factors (low to high accuracy).

CNN method name	Accuracy	Precision	Recall	F1‐score
VGG19	0.92	0.93	0.92	0.92
MobileNetV2	0.92	0.94	0.93	0.93
NASNetMobile	0.93	0.94	0.93	0.93
InceptionResNetV2	0.93	0.95	0.92	0.93
NASNetLarge	0.94	0.95	0.94	0.94
DenseNet201	0.95	0.95	0.93	0.94
InceptionV3	0.95	0.95	0.94	0.94
ResNet152V2	0.95	0.96	0.95	0.95
Xception	0.96	0.96	0.95	0.95
Proposed model	0.98	0.99	0.98	0.98

The class‐wise performance metrics of the proposed model were further analyzed to evaluate its robustness and classification accuracy across different leaf diseases. Figure 9 displays the precision, recall, and F1‐score for each class, offering a detailed insight into the model's effectiveness and statistical robustness.

The model demonstrated outstanding precision, achieving a perfect score of 1.00 for bacterial diseases and pests and infestations. Similarly, pests and infestations and fungal diseases achieved high precision scores of 0.98 and 0.99, respectively. In terms of recall, the model maintained consistent performance across all classes, with pests and infestations and fungal diseases achieving near‐perfect recall values of 1.00 and 0.99, respectively, while bacterial diseases and healthy leaves obtained recall scores of 0.97.

The F1‐score, as the harmonic mean of precision and recall, further highlights the model's robustness, with values ranging from 0.97 to 1.00 across all categories. This performance demonstrates the model's capability to handle class imbalances effectively while maintaining consistent and reliable results across different disease categories. The high precision, recall, and F1‐scores collectively underscore the model's ability to accurately detect various types of diseases and distinguish healthy leaves from diseased ones with exceptional accuracy, showcasing its effectiveness in real‐world applications.

These results underscore the model's reliability in practical applications, showcasing its unparalleled potential for precise classification in real‐world scenarios involving leaf disease detection. The proposed pre‐rained Xception based model significantly outperforms existing benchmarks, as evidenced by its superior statistical metrics, and establishes a new standard for accuracy and reliability in disease detection using advanced deep learning techniques. Through an in‐depth analysis utilizing statistical testing outcomes such as confusion matrices, accuracy and loss metrics, precision and recall curves, and class‐wise performance, we highlight the model's exceptional performance and robustness. This study represents a major step forward in the application of deep learning for plant disease detection, providing a reliable and efficient tool for accurate leaf disease classification.

To improve the interpretability of the model's predictions, we utilized XAI techniques, such as LayerCAM and Saliency Maps. These methods create visual heatmaps that highlight the specific regions of the leaf image that played a significant role in the model's decision‐making process. For example, LayerCAM focused on the specific diseased spots in rose and mango leaf images, while saliency maps highlighted the general areas of infestation.

Figure 10 shows images alongside their explanations, displaying the original image, LayerCAM output, and saliency map prediction for each sample, along with the predicted and actual labels for comparison. These visualization methods enhance our understanding of the model's decision‐making process and its transparency.

FIGURE 10.

Instances of the original image (left), LayerCAM (middle), and saliency maps (right) applied to leaf images with diseases from all classes.

Overall, the statistical analysis indicate that the Xception‐based model accurately categorizes images, significantly reducing the mispredictions between similar classes. The model effectively distinguished between various disease patterns and healthy leaves, demonstrating a robust performance across different scenarios. These visualizations increased stakeholder trust by offering transparency in model decisions, making it suitable for real‐world agricultural applications.

4. Discussion

This study assessed a deep learning framework built upon a modified Xception architecture, leveraging transfer learning for the automated classification of plant diseases and pests. The dataset consisting of Rose, Mango, and Tomato leaves was classified into four major categories: bacterial diseases, fungal diseases, pests and infestations, and healthy leaves, with the model achieving an overall accuracy of 98%. The performance demonstrates the model's potential for supporting sustainable agricultural practices while minimizing the environmental impact of pest management. The results reflect a strong and consistent performance across disease categories, particularly for bacterial diseases with a precision of 1.00 and fungal diseases with 0.98 precision and 0.99 recall. These findings highlight the fairness and accountability of the model in ensuring equitable detection across various disease types. Minor misclassifications of less‐represented pests, such as Spider Mites, were observed initially, which mitigated by upgrading the dataset in training, validation and testing phases to account for more diverse pest characteristics. Though, these misclassifications are relatively insignificant when considering the broader application of the model across different crops and various environments. Compared to previous works, where misclassification for less‐represented categories was higher, the proposed system demonstrates considerable improvement proved by individual class testing statistics in Figure 9. Furthermore, the integration of XAI techniques like LayerCAM and saliency maps enhances transparency in the model's decision‐making process, enabling stakeholders to gain a clearer understanding of its predictions and build trust in AI‐powered agricultural tools.

This research ensures fairness by using a balanced dataset, representing all major disease classes equally. We also employed XAI techniques to provide interpretability and transparency, ensuring that agricultural stakeholders can trust the model's decisions, reducing the risk of biases in underrepresented categories. Our proposed model successfully classified plant conditions with a 98% accuracy rate while maintaining ethical considerations, such as transparency and fairness. We have made a concerted effort to develop a computationally efficient approach by carefully selecting input resolution and optimizing model architecture. Leveraging depthwise separable convolutions, the proposed model significantly reduces computational demands by splitting spatial and channel‐wise operations while preserving accuracy. This architecture is ideal for edge deployment, achieving an optimal balance between performance and scalability. However, while further reducing the model's complexity or resolution could enhance energy and computational efficiency, it would risk compromising accuracy as a critical factor for reliable and precise disease detection. By striking this balance, the study underscores its potential for real‐world scalability and effective implementation in edge computing environments, particularly in resource‐constrained agricultural settings.

By integrating XAI techniques, the system offers not only accuracy but also insights into its decision‐making process, essential for its broader adoption in agricultural settings. The wider implications of this study highlight the potential of AI tools to reduce unnecessary pesticide use, promote sustainable farming practices, and contribute to ensuring global food security. It also has the opportunity to be adapted for real‐time optimization and integration into edge computing environments. Deploying model on edge devices could enable on‐site disease detection, a capability particularly advantageous for resource‐limited agricultural regions. This approach offers comparatively low‐latency solutions while reducing reliance on cloud‐based infrastructures. Future studies should explore these applications to further enhance the system's utility in practical, resource‐constrained settings. Expanding the dataset with images from other geographic regions and plant species will further improve the approach for more specific disease detection. Additionally, integrating RAI practices ensures the model performs ethically, focusing on fairness and avoiding biases in disease detection across different plant species.

Author Contributions

The research was conceptualized by Chowdhury, M.S., Sultan, T., and Jahan, N., with methodology and software contributions from Chowdhury, M.S., and Sultan, T. Validation and formal analysis were conducted by Mridha, M.F., Alfarhood, S., and Safran, M., while investigation and data curation were handled by Chowdhury, M.S., and Sultan, T. Sultan, T., and Jahan, N. prepared the original draft, with review and editing by Chowdhury, M.S., Alfarhood, S., and Che, D. Visualization was managed by Sultan, T., and Safran, M., supervision by Mridha, M.F., project administration by Safran, M., and Che, D., and funding acquisition by Alfarhood, S. All authors approved the final manuscript.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Data S1 Peer review.

Data Availability Statement

The datasets generated or analyzed during this study are publicly available and can be accessed at the provided link.