A review of deep learning architectures for plant disease detection

TL;DR

This review summarizes how deep learning methods, like CNNs and object detectors, are used to detect plant diseases through images, highlighting challenges and solutions for real-world applications.

Contribution

The paper systematically reviews DL architectures and strategies for plant disease detection, emphasizing interpretability and real-time deployment solutions.

Findings

CNN architectures like VGG, ResNet, and EfficientNet are effective for plant disease classification.

XAI techniques such as Grad-CAM improve model interpretability in plant pathology.

Edge computing and synthetic data generation help address computational and data challenges.

Abstract

The rapid advancement of deep learning (DL) has revolutionized plant disease detection by enabling highly accurate, image-based diagnostic solutions. This review provides a comprehensive synthesis of DL-based methodologies for plant disease detection, systematically structured around the key stages of the modeling pipeline, encompassing data acquisition, preprocessing, augmentation, classification, detection, segmentation, and deployment.

The review focuses on evaluating convolutional neural network (CNN) architectures such as VGG, ResNet, EfficientNet, and DenseNet across diverse experimental contexts. Classification strategies are categorized according to their integration of visualization techniques (e.g., saliency maps, Grad-CAM) to enhance model interpretability, emphasizing the pivotal role of explainable artificial intelligence (XAI) in plant pathology. Object detection models are systematically examined within both one-stage (YOLO, SSD) and two-stage (Faster R-CNN) paradigms. Furthermore, critical challenges—such as environmental variability, data imbalance, and computational constraints—along with potential solutions including transfer learning, synthetic data generation using generative adversarial networks (GANs) and diffusion models, and edge computing for real-time deployment, are comprehensively discussed.

This review summarizes best practices for dataset selection and model optimization for mobile platforms, emphasizing their role in improving the efficiency and accuracy of plant disease detection systems.

Deep learning-based methods show strong potential to enhance precision and resilience in real-world plant disease detection and monitoring.

Figures

Equations

Full text

Introduction

Population growth has driven a steady increase in the demand for various plant products. Various crop protection strategies have been employed for decades to combat plant diseases, playing a crucial role in meeting the increasing demand for large quantities of plant-based food products (Kaya and Gursoy, 2023). However, the rising costs associated with pest infestations and plant diseases pose an increasingly serious threat to crop yields, thereby jeopardizing harvest success. According to the Food and Agriculture Organization (FAO), plant diseases cause approximately 20%–40% of global crop losses each year, posing a major threat to food security and sustainable agriculture (FAO, 2021). Furthermore, biotic stresses represent a major threat to agricultural yields, leading to considerable losses in food production. Plants play vital roles in society by providing economic and social benefits, supporting environmental protection, advancing agricultural development, and ensuring food supply. However, plant-related disciplines, including species identification, early disease detection, and yield estimation, often require specialized expertise and considerable human effort (Kaya and Gursoy, 2023; Cetinkaya and Tandirovic Gursel, 2025).

Plant diseases represent a major challenge in agriculture, exerting a substantial impact on crop production. They contribute substantially to reduced crop yields and adversely affect both the operational efficiency and financial performance of farms. They also reduce productivity and pose significant obstacles to the smooth operation of farming activities. The impact of plant diseases is extensive, encompassing issues from farm-level productivity losses to global food safety concerns. Technical expertise and effective management of plant diseases are therefore essential for achieving sustainable agriculture.

Plant diseases are primarily caused by diverse bacteria, fungi, viruses, and other naturally occurring pathogens throughout the plant life cycle. Several methods have been developed for identifying and classifying different leaf diseases. The most common technique is direct visual inspection with the naked eye; however, it is not always sufficiently effective. This approach can be time-consuming and demands prompt decision-making to prevent the spread of disease to healthy plants. Owing to these limitations, researchers increasingly rely on machine learning (ML) and deep learning (DL) methods for plant disease detection and classification. Plant disease detection has emerged as a commercially viable research domain integrating image processing, ML, DL, and computer vision techniques. These advanced technologies facilitate efficient and timely disease detection, resulting in healthier plants and improved crop yields.

1.1. Challenges

1.1.1. Environmental factors

Environmental factors play a critical role in the onset and spread of plant diseases. Disease symptoms are influenced by temperature, humidity, and soil conditions, which complicates the development of reliable detection models. Climate change profoundly affects various stages of a plant’s life cycle, with temperature fluctuations influencing key phenological events such as flowering and leaf emergence (Keller and Shea, 2021). Changes in these ecological variables can disrupt species synchrony, resulting in ecological imbalance.

To address these challenges, Isinkaye et al. (2024) proposed a novel approach combining DL with content-based filtering to enhance plant disease detection and treatment recommendations. Their study highlights the influence of environmental conditions on the variability of disease symptoms and detection accuracy. Similarly, Corlouer et al. (2024) introduced the concept of ecological classification, integrating environmental data with quantitative genetics to improve the performance of plant disease detection models.

In addition, Lee and Yun (2023) developed a model incorporating historical environmental data—including temperature, humidity, and soil parameters—to more accurately assess plant disease risk. Their results emphasize the importance of incorporating environmental parameters into disease prediction models to enhance model robustness and practical applicability.

1.1.2. Symptom variability

The variability of plant disease symptoms presents a major challenge for DL-based plant disease detection. The same plant disease may manifest differently depending on various environmental and biological factors. Several strategies have been proposed to address symptom variability. A large dataset containing images of diseased plants exhibiting diverse symptoms is employed as one approach in plant disease detection studies. This enables the model to recognize multiple types of plant diseases more effectively (Ferentinos, 2018).

Another approach involves the use of a technique known as transfer learning. In transfer learning, a model is pretrained on a large dataset of images unrelated to the target domain, such as natural or medical images (Dinc and Kaya, 2024; Tandirovic Gursel and Kaya, 2025). The model is subsequently fine-tuned using a smaller dataset of diseased plant images. This approach helps the model learn generalizable features relevant to plant disease detection and enhances its robustness to symptom variability (Brahimi et al., 2018; Arnal Barbedo, 2019).

Researchers have also developed novel DL architectures specifically designed for plant disease identification. These architectures are designed to learn complex relationships among multiple features within an image. This enables more accurate identification of plant diseases, even when symptom expressions vary (Lu et al., 2022). These studies highlight the importance of employing large datasets, transfer learning, and advanced DL architectures to improve the reliability of plant disease detection models despite symptom variability.

1.1.3. Image processing challenges

Image processing is a fundamental component of plant disease detection; however, several challenges limit its efficiency. These challenges include image noise, blurring, inconsistent data acquisition conditions, and the inherent complexity of feature extraction.

Lighting variability and limited imaging equipment commonly contribute to noise and blurring, which can obscure critical visual details required for accurate disease identification. Deep architectures such as deep belief networks and deep Boltzmann machines have been developed to address these challenges by effectively denoising images and extracting robust features to ensure high classification accuracy for diseased plant images (Shoaib et al., 2023).

Collecting high-quality agricultural images poses several challenges, including variable weather conditions, uneven illumination, and diverse camera perspectives. These factors contribute to image quality variability, complicating the accuracy of disease detection. The use of unmanned aerial vehicles (UAVs) for image acquisition introduces additional limitations, including weather dependency and restricted flight duration (Neupane and Baysal-Gurel, 2021).

Filtering techniques such as box, Gaussian, gradient, and Laplacian filters are employed to mitigate noise-related issues. These filters enhance image quality by smoothing noise while preserving critical structural features. They are often combined with deep denoising autoencoders (DDA) to remove noise from plant leaf images and improve the accuracy of subsequent disease detection models (Shoaib et al., 2023).

Extracting discriminative features from plant images is challenging due to the diverse visual manifestations of diseases and complex natural backgrounds. Advanced feature extraction techniques such as thresholding, edge detection, template matching, Hough transform, and histogram of oriented gradients (HoG) are used to identify diseased regions. The irregular occurrence of diseases and heterogeneity of symptoms necessitate efficient feature extraction techniques (Sarkar et al., 2023).

1.2. Objectives

This study aims to provide a comprehensive overview of research on plant disease detection and to explain the various methodological approaches employed by researchers in the literature. The review covers several key stages, including image preprocessing, segmentation, feature extraction, and classification. A comprehensive analysis of practical DL algorithms applied to plant disease diagnosis is also presented. This review specifically focuses on studies conducted between 2018 and 2023. To illustrate the quality of research during this period, we present a list of SCI- and Scopus-indexed publications along with their key contributions. This process provides an overview of the current literature and serves as a practical guide for newcomers to the field of plant disease detection. The review is intended as a reference tool to facilitate clear understanding of the methods and models used in this research domain.

1.3. Contributions

This paper presents an overview of plant disease detection methods developed between 2018 and 2024. To facilitate the work of researchers, the key aspects are summarized as follows:

The review outlines the advantages and limitations of employing DL methods for plant disease detection.Particular emphasis is placed on algorithms commonly applied in this context to distinguish between ML and DL models.A comprehensive set of tables lists all datasets used in the reviewed studies and provides relevant details for each.The overview illustrates the distribution of selected studies across multiple sectors, including universities, governmental institutions, private organizations, industries, and academia–industry collaborations.The review assists researchers in selecting appropriate methods and datasets according to their research objectives. This review serves as a valuable resource for researchers seeking an overview of leaf disease detection methods conducted within the specified timeframe.

1.4. Search process

To ensure a comprehensive and systematic literature search on DL-based plant disease detection, a well-defined search process is essential. This process involves several essential steps to collect relevant scientific articles and datasets.

1.4.1. Database selection

To identify relevant literature, several well-established academic databases were utilized, including:

IEEE Xplore—for artificial intelligence (AI) and computer vision applications in agriculture.ScienceDirect—for research in agricultural and biological sciences.**SpringerLink—**for multidisciplinary studies involving deep learning applications.Google Scholar—for a broad range of related scholarly articles.PubMed—for bioinformatics-related research on plant diseases.

1.4.2. Search keywords and query design

Search queries were systematically formulated to retrieve high-quality and peer-reviewed publications. The following keywords and Boolean operators were employed:

(“Deep Learning” OR “Convolutional Neural Network” OR “Machine Learning”) AND (“Plant Disease” OR “Leaf Disease”) AND (“Detection” OR “Classification”)(“Image Processing” AND “Plant Health Monitoring”)(“Hyperspectral Imaging” OR “Multispectral Imaging”) AND (“Disease Diagnosis”)

1.4.3. Inclusion and exclusion criteria

To ensure the relevance and quality of the selected studies, specific inclusion and exclusion criteria were applied.

Inclusion

Peer-reviewed journal articles.Studies published from 2018 onwards.Experimental studies focusing on real-world plant disease detection.Articles providing open-source or locally collected datasets.

Exclusion

Studies lacking experimental validation.Studies employing outdated methodologies.Review papers that do not provide novel insights or findings.Preprints hosted on servers such as arXiv and bioRxiv were excluded.

1.4.4. Data extraction and categorization

The extracted studies were categorized according to the following criteria:

**Deep learning architecture used—**for example, convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformer-based models.Type of dataset—including PlantVillage, real-world agricultural field images, and synthetic datasets.Evaluation metrics—such as accuracy, precision, recall, F1-score, AUC-ROC, and mean average precision (mAP).

1.4.5. Verification and cross-referencing

Cross-referencing was employed to verify citations in highly cited studies, ensuring that no significant work was overlooked. This systematic search process ensures a comprehensive and up-to-date literature review, establishing a robust foundation for analyzing recent advances in DL-based plant disease detection.

Deep learning

The concept of DL was first introduced by Hinton et al. (2006). However, the foundation of deep neural networks (DNNs) was established much earlier, with notable contributions from Lecun et al. (1998) on convolutional neural networks (CNNs) and Goodfellow et al. (2016) on the general principles of deep learning (DL). DL utilizes DNNs for data analysis and hierarchical feature learning. Each layer functions as a processing unit, extracting low-level features and integrating them into higher-level representations. This hierarchical feature extraction enables DL models to automatically learn complex patterns from raw data. This approach overcomes the limitations of conventional algorithms that depend heavily on manually engineered features.

Traditional image classification and recognition methods often rely on manually designed features (Topuz and Kaya, 2025), limiting their ability to identify complex patterns. DL overcomes these limitations through unsupervised learning directly from raw images, enabling the extraction of multilevel features that capture low-, mid-, and high-level semantic information (Kamilaris and Prenafeta-Boldú, 2018; Hamroun et al., 2025; Kaya et al., 2025). This capability makes DL a robust framework for data analysis and pattern recognition.

Traditional approaches to detecting plant diseases and pests have relied on image recognition techniques based on manually designed features. These approaches are often complex, requiring extensive expertise, and their performance frequently depends on subjective interpretation. DL provides an effective alternative by autonomously learning features from large-scale datasets without manual intervention. Multilayered models possess continuous learning and feature extraction capabilities, facilitating image classification and recognition in plant pathology. As the volume of training data and computational power increases, the representational capacity of DNNs correspondingly improves.

In recent years, DL-based approaches, particularly CNN models, have gained prominence in plant disease and pest image recognition, providing more accurate and efficient solutions than traditional methods. As illustrated in Figure 1, a CNN model typically consists of several layers: an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. During a single iteration, convolutional and pooling layers perform a sequence of alternating operations. This process involves multiple alternating cycles between convolutional and pooling layers, producing a cascading effect. A distinctive property of CNN architecture is that neurons in convolutional and pooling layers are not fully connected.

The convolutional layer plays a key role in extracting salient features from input data through convolutional operations. These convolutional operations apply filters to detect spatial patterns and relationships within the input data. Subsequently, pooling layers reduce the spatial dimensions, decreasing computational cost while preserving critical feature information. The alternating structure of convolutional and pooling layers enhances the network’s capacity to capture complex hierarchical patterns in data (Barbedo, 2018). The connections between neurons in the convolutional and pooling layers are partial rather than fully connected, enhancing model efficiency. The connectivity paradigm increases network efficiency by selectively retaining essential information and discarding redundant details.

Consider the visualization of leaf images in a CNN framework applied for plant disease identification. Figure 2 illustrates the progressive transformation of leaf images at each processing stage. This visual representation demonstrates how a CNN systematically extracts features and facilitates accurate classification in complex problem domains such as plant disease identification.

2.1. Architectures and tools

DL is characterized by the availability of well-established benchmark architectures that provide researchers with a solid foundation for model development, eliminating the need to build models from scratch. These architectures include AlexNet (Krizhevsky et al., 2012), CaffeNet (Jia et al., 2014), VGG (Simonyan and Zisserman, 2015), GoogleNet (Szegedy et al., 2015), and Inception-ResNet (Szegedy et al., 2017), among others, each offering distinct design principles and feature extraction capabilities. Furthermore, Canziani et al. (2017) evaluated and selected architectures based on their specific strengths and suitability for targeted applications and datasets. By pretraining on large-scale datasets such as ImageNet (Deng et al., 2009), neural network architectures acquire pretrained weights that facilitate feature learning and accelerate model convergence. Consequently, pretrained models can deliver accurate classifications across specific problem domains (Pan and Yang, 2010).

Researchers employing DL techniques utilize a variety of tools and platforms. The most widely used frameworks include TensorFlow, Keras, Theano, PyTorch, Caffe, Pylearn2, TFLearn, and the Deep Learning Toolbox in MATLAB. These frameworks provide a flexible environment for experimenting with various architectures and training models on diverse datasets. The individual functionalities of these open-source frameworks are summarized in Table 1. Theano and Caffe simplify implementation by integrating popular architectures such as AlexNet, VGG, and GoogleNet as built-in libraries or predefined classes. PyTorch and TensorFlow demonstrate high scalability through integration with third-party libraries and diverse network architectures, offering remarkable efficiency in training large-scale networks on graphics processing units (GPUs) and enabling accelerated training performance.

Applications of deep learning in plant disease detection

This section provides a comprehensive overview of methods employed to identify plant diseases and pests through DL applications. Figure 3 visually represents the major components of DL-based systems, including classification, detection, and segmentation. Additionally, the following subsections are organized according to the specific DL methods employed.

3.1. Feature extraction and selection

In the early stages of DL-based plant disease and pest classification research, researchers primarily relied on the inherent feature extraction capabilities of CNN models. These CNN-based feature extraction methods were combined with traditional classifiers to enhance overall classification performance (Singh et al., 2018). Input images are fed into a pretrained CNN to extract representative image features, which are subsequently passed to a conventional ML classifier for final classification.

Yalcin and Razavi (2016) proposed a CNN-based framework for image feature extraction, integrating support vector machine (SVM) classifiers with various kernels and feature descriptors such as local binary patterns (LBP) and global image descriptor (GIST). Similarly, Mohameth et al. (2020) developed CNN architectures incorporating transfer learning and deep feature extraction using the publicly available PlantVillage dataset. Their findings identified the SVM as the most effective classifier for leaf disease detection, demonstrating the potential of DL-based technological solutions in plant pathology.

Hasan et al. (2019) identified and classified nine rice diseases using features extracted from a deep CNN model, which were subsequently processed by a SVM classifier, achieving 97.5% accuracy. Transfer learning is a DL strategy that transfers knowledge acquired from previous tasks to improve learning efficiency on new tasks. In this approach, only specific layers of pretrained networks are fine-tuned with new datasets, substantially reducing the data requirements typically associated with ML tasks (Pan and Yang, 2010).

Sethy et al. (2020) developed deep CNNs for rice leaf disease identification through deep feature extraction. Using 5932 images representing four rice diseases—bacterial blotch, pat, brown spot, and tungro—the study evaluated eleven CNN models with transfer learning, integrating deep features with SVMs. The results indicated that SVM classifiers using deep features outperformed direct transfer learning approaches in terms of accuracy. The study also evaluated lightweight architectures such as MobileNetV2 and ShuffleNet. Evaluation metrics included accuracy, sensitivity, specificity, false positive rate (FPR), F1-score, and training time. The highest performance was achieved using ResNet50-derived features with an SVM classifier, yielding an F1-score of 0.9838.

Ramesh et al. (2018) employed the random forest (RF) algorithm for papaya leaf classification and compared its performance with other classifiers, including logistic regression (LR), classification and regression tree (CART), SVM, k-nearest neighbors (KNN), and naive Bayes. For object recognition, they employed histogram of oriented gradients (HOG) for feature extraction, incorporating color histograms, Haralick texture features, and Hu moments. The resulting feature vector was subsequently used for classification. The results indicated that the RF algorithm achieved the highest accuracy of 70.14%.

Islam et al. (2017) employed a multiclass SVM classifier to analyze potato leaf images, distinguishing between early blight, late blight, and healthy samples. The dataset consisted of 200 diseased and 100 healthy leaf images. The gray-level co-occurrence matrix (GLCM) method was used to extract 10 texture and color features. A linear-kernel SVM was trained on 180 images and tested on 120, with performance evaluated using precision, accuracy, F1-score, and recall metrics. The model achieved 95% accuracy with a 60–40 train–test split, and 93.7% accuracy using five-fold cross-validation. The results demonstrate the effectiveness of multiclass SVM classifiers for potato disease classification.

Recent advances in DL have introduced architectures beyond CNNs, including transformer-based models such as the vision transformer (ViT) and swin transformer (Khan et al., 2023a). These architectures employ self-attention mechanisms instead of convolutional operations, enabling them to better capture the global dependencies inherent in plant disease images. Studies have shown that ViT models perform better when trained on large-scale datasets with high-resolution images, whereas CNNs remain more efficient for smaller datasets (Ogrekci et al., 2023).

Additionally, generative adversarial networks (GANs) have been employed to synthesize artificial images, thereby enhancing the generalization capability of plant disease classification models (Bi and Hu, 2020). Multimodal DL approaches combining hyperspectral and thermal imaging have also been employed to detect early-stage infections before visible symptoms appear (Terentev et al., 2022).

These developments underscore the continuous advancement of DL techniques in plant disease detection. They emphasize the importance of integrating diverse feature extraction techniques, leveraging pretrained models, and adopting alternative network architectures to achieve higher accuracy and robustness in practical agricultural applications.

3.2. Disease classification

Image classification represents a fundamental task within the field of image processing. It has played a crucial role in the advancement of computer vision, with major breakthroughs achieved through the adoption of DL models and novel approaches. Numerous studies have employed various methods to identify and classify plant diseases. This section presents a detailed analysis of plant disease image classification approaches, further elaborated in the following subsections.

3.2.1. Classification with visualization techniques

In the following studies, visualization techniques were integrated with DL methods to enhance interpretability and achieve a deeper understanding of disease characteristics. Brahimi et al. (2018) employed a saliency map-based visualization technique to classify disease symptoms in affected plants. The authors utilized the CaffeNet CNN architecture to identify 13 plant diseases, outperforming existing models for plant disease classification. The proposed model achieved an accuracy of 99.76%. Atila et al. (2021) introduced the EfficientNet architecture and compared its performance with other state-of-the-art models using the PlantVillage dataset. They applied visualization techniques to augment and interpret the dataset more effectively. Training was performed on both the original and augmented datasets using transfer learning, with all layers set as trainable. The model achieved accuracy rates of 99.91% and 99.97% on the original and augmented datasets, respectively. Nagasubramanian et al. (2019) employed a three-dimensional deep convolutional neural network (DCNN) model along with saliency map visualization to distinguish between healthy and infected soybean leaves affected by blight. The classification model achieved an accuracy of 95.73%.

Feng et al. (2021) utilized hyperspectral imaging to rapidly and accurately detect rice diseases across four distinct rice varieties. Because developing a separate classifier for each variety is time- and resource-intensive, the study adopted deep transfer learning to enhance detection performance across rice varieties. A total of three deep transfer learning approaches were evaluated. Among these, fine-tuning exhibited the highest transferability, achieving accuracy rates above 88% on test sets for most transfer tasks. The deep CORAL method also performed well, achieving over 80% accuracy across four tasks and outperforming the deep domain confusion (DDC) approach. Additionally, a multitask transfer learning strategy produced promising results, demonstrating the effectiveness of both pairwise and multitask transfer mechanisms. Saliency maps were employed to visualize the primary wavelength ranges identified by the CNN, both with and without transfer learning, revealing substantial overlap in the captured spectral features.

Gonzalez-Huitron et al. (2021) investigated the training and evaluation of four CNN models developed to identify diseases in tomato leaves. The study employed the PlantVillage dataset, consisting of 18,160 red–green–blue (RGB) images categorized into 10 distinct classes for transfer learning. The selected CNN models incorporated depthwise separable convolutional architectures optimized for deployment on low-power devices. Model performance was assessed both quantitatively and qualitatively using standard evaluation metrics and saliency maps to enhance interpretability. To demonstrate the practical applicability of their method, the authors developed a graphical user interface (GUI) enabling user interaction with the system.

Cao et al. (2022) investigated early detection of rice bacterial leaf blight (BLB) infections using hyperspectral imaging combined with a spectral dilated convolutional 3D CNN (SDC-3DCNN). Hyperspectral images captured from infected rice leaves during the tillering stage were preprocessed using Savitzky–Golay smoothing. Wavelengths between 450 and 950 nm were selected, and features extracted using principal component analysis (PCA) and RF algorithms were input into the SDC-3DCNN model. The model’s performance was evaluated using different input features and spectral dilation ratios.

Furthermore, saliency maps were used to visualize the sensitivity of individual wavelengths, providing insight into the model’s focus regions and interpretability. Experimental results indicated that the SDC-3DCNN model achieved 95.44% accuracy when using 50 characteristic wavelengths identified by the RF algorithm with a dilation ratio of 5. Saliency analysis identified sensitive wavelength bands within the 530–570 nm range, corresponding to the critical wavelengths extracted by the RF algorithm. These findings underscore the effectiveness of integrating hyperspectral imaging with DL for early detection of asymptomatic rice BLB infections, offering critical support for early warning systems and preventive measures in rice cultivation.

Anim-Ayeko et al. (2023) proposed a ResNet-9 model for classifying late blight in potato and tomato leaf images. The PlantVillage dataset, initially containing 3990 training samples, was augmented to improve model generalization capability. After data augmentation, an extensive hyperparameter optimization process was performed to determine the optimal configuration for model training. The optimized ResNet-9 model was subsequently trained and evaluated on a test set comprising 1331 images. Experimental results demonstrated high classification performance, achieving 99.25% test accuracy, 99.67% precision, 99.33% recall, and an F1-score of 99.33%.

Sun et al. (2019) proposed an image analysis approach for detecting leaf diseases in tea plants, involving superpixel block generation, key-point detection, and extraction of blurred contours of salient regions. Texture features were removed, and a classification map was generated using a SVM classifier. Morphological and algebraic operations were applied for block restoration, resulting in a highly accurate saliency map. The method was evaluated on 261 diseased leaf images, achieving 98.5% accuracy, 96.8% precision, 98.6% recall, and a 97.7% F1-score.

In a related study, Tembhurne et al. (2023) developed a classification model based on the MobileNet architecture, augmented with additional hidden layers and fine-tuned using saliency map-based optimization. The model was trained on a dataset containing 12,318 images and designed to classify inputs into 64 distinct categories across 22 plant species. Experimental results showed that the proposed model achieved 95.94% classification accuracy, underscoring its effectiveness for multiclass plant disease detection.

Khan et al. (2020) outlined a five-step process for classifying cucumber leaf diseases, consisting of infected spot segmentation, image enhancement, feature selection, deep feature extraction, and classification. Image enhancement was applied to boost local contrast, making infected regions more distinguishable. For segmentation, a novel approach combining Sharif’s saliency method with active contour segmentation was employed. Pretrained models, specifically VGG-19 and VGG-M, were used for feature extraction, with the most relevant features selected based on local entropy and standard deviation. The resulting features were subsequently fed into a multiclass support vector machine (SVM) classifier to identify the disease, achieving a classification accuracy of 98.08%.

Additionally, some researchers have employed heatmap visualization, a common DL interpretability tool that helps analyze neural network outputs, particularly in prediction and segmentation tasks. These visualizations assist analysts and researchers by identifying regions of interest that the model focuses on. Wiesner-Hanks et al. (2019) applied a heatmap-based visualization approach to accurately outline maize disease lesion contours. Using images captured by unmanned aerial vehicles (UAVs), their model achieved exceptional accuracy in classifying lesions, even at the millimeter scale. The achieved accuracy rate of 99.79% represents the highest reported performance for aerial plant disease classification to date. Albattah et al. (2022) proposed an improved CenterNet framework integrated with a DenseNet-77 backbone. The proposed framework consisted of three main steps: creating region-of-interest (ROI) annotations, applying the enhanced CenterNet with DenseNet-77 for feature extraction, and using CenterNet for plant disease detection and classification.

DeChant et al. (2017) introduced an innovative computational pipeline utilizing convolutional neural network (CNN) models to address challenges associated with limited data and irregularities in field-grown plant images. The approach involved training multiple CNNs to classify small image patches displaying northern leaf blight (NLB) lesions. Predictions from these networks were aggregated into heatmaps, which were then used as input to train a final CNN that classified entire images as diseased or healthy. The system achieved a classification accuracy of 96.7% on a test dataset comprising images not used during training.

Nawaz et al. (2024) proposed a CenterNet-based methodology integrating a ResNet50 backbone with a spatial channel attention mechanism to extract deep, disease-specific features from image samples. This enhanced feature extraction process was implemented within the single-stage detection architecture of the CenterNet framework. The developed model, named CoffeeNet, was evaluated using the Arabica coffee leaf dataset, which contains images captured under realistic and complex environmental conditions. Experimental results showed that CoffeeNet achieved 98.54% classification accuracy and a mean average precision (mAP) of 0.97, demonstrating strong robustness in real-world applications.

In a related study, Kumar et al. (2022) developed an automated detection system for three major corn diseases: common rust, Cercospora spot, and northern leaf blight (NLB). The system combined image recognition techniques with DL methods, specifically employing Faster R-CNN with a ResNet50 backbone network. The model achieved 93.5% classification accuracy when evaluated on real-time imaging data. Additionally, Chowdhury et al. (2021) employed the EfficientNet architecture to classify 18,161 tomato leaf images using a modified U-Net framework. This model achieved 98.66% accuracy, with the EfficientNetB7 variant outperforming others—reaching 99.95% in binary classification and 99.12% in six-class classification tasks using segmented images.

In cases where certain diseases manifest in early stages without visible symptoms, indicative signs may appear in regions of the electromagnetic spectrum invisible to the human eye. In such scenarios, hyperspectral and multispectral imaging techniques prove highly beneficial. For instance, advanced imaging techniques enable the detection and analysis of subtle variations in electromagnetic wavelengths, offering insights that would otherwise remain undetected through conventional visual observation.

Park et al. (2018) proposed a hyperspectral image analysis method utilizing the minimum redundancy maximum relevance (mRMR) algorithm to select the most informative spectral bands. By integrating a CNN with a fully connected network (FCN), the proposed deep neural architecture effectively classified dimensionally reduced hyperspectral data. The approach successfully identified five key spectral bands, achieving higher classification accuracy than conventional RGB imaging. In a related study, Polder et al. (2019) employed hyperspectral imaging to detect Potato virus Y (PVY) infections in seed potatoes. Hyperspectral images were acquired during field experiments using a line-scan hyperspectral camera with a spatial resolution of 5 mm. The study utilized two CNNs for model training, using data collected from two independent field experiment series.

Nguyen et al. (2021) applied hyperspectral imaging at the plant level to detect and classify grapevines infected with the grapevine vein clearing virus. The experiments involved both pixel-by-pixel and image-level classification approaches. The methodology combined DL and ML architectures for analysis. The RF classifier exhibited superior performance in both pixel-level and image-level classification, particularly in scenarios involving high-dimensional feature spaces. Xu et al. (2023) proposed residual feedback ensemble-CNN (RFE-CNN), an integrated DL framework combining residual channel attention blocks (RCAB), feedback blocks (FB), enhanced memory layers (EML), and CNNs. A pair of parallel CNNs was used for feature extraction, optimized via RCAB, iteratively trained using FB, and subsequently processed with CNN and EML for final classification. RFE-CNN outperformed benchmark models, demonstrating superior time efficiency, recognition accuracy, and adaptability, achieving an overall accuracy of 98.83%.

Yong et al. (2023) focused on automated detection of basal stem rot (BSR) at the seedling stage using pretrained DL models and hyperspectral imaging. Aerial images of oil palm seedlings were segmented into three distinct regions to assess spectral differences across various leaf positions. To evaluate the impact of background imagery on recognition accuracy, seedling images were automatically segmented using a region-based CNN (R-CNN) incorporating mask region proposals. Gao et al. (2023) proposed a multidimensional CNN framework integrating spectral, spatial, and spectral–spatial features for potato disease detection. The framework integrates a 1D-CNN for spectral feature extraction, a 2D-CNN for spatial representation learning, and a 3D-CNN for spectral–spatial data analysis. Convolution operations in the 1D-CNN and 2D-CNN were optimized to minimize data loss during processing. Experimental evaluations on real potato disease datasets achieved a detection accuracy of up to 99.87%, confirming the model’s effectiveness.

These visualization techniques form integral components of the broader discipline of explainable artificial intelligence (XAI)—a rapidly evolving field aimed at improving the interpretability, transparency, and reliability of complex ML models, particularly those employing DL architectures. In agriculture, the implementation of XAI methodologies plays a pivotal role in enhancing the reliability of automated decision-making processes, including disease diagnosis and crop management. By allowing agronomists and stakeholders to better understand how specific features or patterns influence model predictions, XAI facilitates improved validation and broader acceptance of these technologies. Moreover, visualization-based classification methods constitute a key subset of XAI strategies specifically designed for plant disease detection, where visual interpretability is essential. These techniques not only assist in localizing and identifying diseased regions within plant imagery but also promote transparent interaction between algorithms and end-users, thereby increasing confidence in automated agricultural diagnostics.

Figure 4 provides a comparative summary of classification accuracies achieved various DL models utilizing visualization techniques. These methods include saliency maps, heatmaps, hyperspectral imaging, and spectral attention mechanisms, each applied to distinct datasets. The figure highlights the superior performance of EfficientNet-, DenseNet-, and CenterNet-based architectures in achieving peak accuracy, with several models exceeding the 99% threshold.

3.2.2. Classification without visualization techniques

Applying DL techniques for plant disease classification without visualization methods represents a novel direction in agricultural diagnostics. Traditional diagnostic methods often rely on visual inspection of symptoms, which are constrained by the observer’s expertise and environmental conditions. In contrast, DL models can be trained on diverse datasets that include genetic and phenotypic information related to plant diseases. This enables models to learn complex patterns and associations, allowing disease identification based on nonvisual features. This approach provides a more comprehensive and precise means of diagnosing diseases, reducing dependence on visual cues and proving particularly valuable when direct observation is limited.

Naik et al. (2022) conducted a detailed study to classify five chili diseases using 12 CNN models. The analysis included 12 models, namely EfficientNetB0, InceptionV3, AlexNet, DarkNet53, ShuffleNet, SqueezeNet, DenseNet201, ResNet101, VGG19, MobileNetV2, NasNetLarge, and XceptionNet. The VGG19 model demonstrated the highest performance, achieving 83.54% accuracy without data augmentation or visualization techniques.

Dai et al. (2024) proposed a multilevel deep feature fusion network, termed deep feature network with pyramidal squeezed attention network (DFN-PSAN) for plant disease classification. The architecture integrates the YOLOv5 backbone and neck components, leveraging their strengths in feature extraction. The model further incorporates pyramidal squeezed attention (PSA) mechanisms and multiple convolutional layers to enhance the overall performance of the PSAN framework. This network effectively fuses and processes multilevel deep features from the DFN, while the PSA module provides pixel-level attention to highlight disease-related regions in plant images. Experiments on three plant disease datasets showed that the DFN-PSAN framework achieved an average accuracy and F1-score of 95.27%, underscoring its effectiveness in accurate plant disease diagnosis.

Dey et al. (2022) compared the performance of pretrained deep CNN models for classifying rice diseases such as Hispa, brown spot, leaf blight, and Nitrogen, Phosphorus, and Potassium (NPK) deficiency symptoms. Training the DL models on different combinations of public datasets revealed that mixed datasets yielded the best results, with the VGG19 model achieving 91.8% accuracy. The study demonstrated that a simple, well-structured CNN outperformed more complex models in classifying phosphate-deficient leaves.

Vallabhajosyula et al. (2022) introduced an automated plant disease classification framework, termed deep ensemble neural networks (DENN), which leverages transfer learning using pretrained models. To reduce overfitting and enhance model robustness, various data augmentation techniques—including image enhancement, rotation, scaling, and translation—were employed. The proposed framework was evaluated using the PlantVillage dataset, which includes 38 classes spanning 14 crop species. Faisal et al. (2023) aimed to improve classification accuracy for citrus plant datasets by employing state-of-the-art transfer learning architectures. Their methodology incorporated convolutional neural networks (CNNs) with pretrained architectures, including EfficientNetB3, ResNet50, MobileNetV2, and InceptionV3, focusing primarily on disease detection and categorization in citrus plants.

Alirezazadeh et al. (2023) examined the impact of integrating the convolutional block attention module (CBAM) on improving plant disease classification performance in CNN architectures. CBAM—a lightweight attention mechanism—was embedded into well-established CNN architectures. The models were fine-tuned on the DiaMOS Plant dataset, and experimental results showed that EfficientNetB0 combined with CBAM achieved the highest classification accuracy (86.89%), outperforming the standalone EfficientNetB0 model. Andrew et al. (2022) proposed the use of pretrained CNN models for plant disease classification, utilizing the PlantVillage dataset. Among the evaluated models, DenseNet121 achieved superior performance, with a classification accuracy of 99.81%.

Amin et al. (2022) proposed a model employing pretrained CNNs, specifically DenseNet121 and EfficientNetB0. These models were used to extract features from corn leaf images. Data augmentation techniques expanded the training dataset by introducing image variations, enabling the model to capture complex patterns. The proposed model achieved a classification accuracy of 98.56%, outperforming ResNet152 and InceptionV3, which require higher computational resources. Specifically, the proposed model outperformed ResNet152 and InceptionV3, which achieved accuracies of 98.37% and 96.26%, respectively.

Coulibaly et al. (2019) introduced a transfer learning-based approach for identifying mildew diseases in pearl millet. They employed the classical VGG16 CNN model, pretrained on the publicly available ImageNet dataset. The experimental results demonstrated strong performance, with an accuracy of 95% and a recall of 94.5%. In another study employing VGG architectures, Oppenheim et al. (2019) used potato images captured under natural lighting conditions, encompassing diverse sizes, shades, and shapes. The classification of infected potato images was achieved through fine-tuning of the VGG network. Experimental results showed that both transfer learning and training from scratch produced successful classification outcomes.

Fan et al. (2022) proposed a methodology integrating transfer learning-based deep feature extraction, traditional handcrafted features, and center loss to enhance discriminative capability. Experimental evaluations on three datasets—two apple leaf and one coffee leaf—achieved classification accuracies of 99.79%, 92.59%, and 97.12%, respectively. Fraiwan et al. (2022) employed the InceptionResNetV2 convolutional neural network (CNN) combined with transfer learning to identify rice leaf diseases. After parameter optimization tailored to the classification task, the proposed model achieved an accuracy of 95.67%.

Rancon et al. (2019) employed transfer learning-based feature extraction techniques to classify vine plant diseases. The research was conducted across two vineyards in the Bordeaux region of France (Aquitaine), where images of healthy and diseased vine plants were collected during the summer of 2017. The collected images were meticulously annotated at the leaf level, resulting in a dataset of approximately 6000 images (224 × 224 pixels) categorized into red and white cultivars. For classification, the authors compared the efficacy of scale-invariant feature transform (SIFT) encoding with that of pretrained DL feature extractors. The highest overall accuracy of 91% was achieved using MobileNet feature representations.

Ahmad et al. (2021) introduced a sequential transfer learning methodology designed to accelerate convergence, mitigate overfitting, and prevent negative transfer across diverse domains. The proposed system was evaluated on two distinct plant disease datasets: the PlantVillage dataset and a more challenging pepper disease dataset. Results showed classification accuracies of 99% on the pepper dataset and 99.69% on the PlantVillage dataset, demonstrating the approach’s robustness across domain contexts. Abbas et al. (2021) developed a DL framework for tomato disease detection, leveraging a conditional generative adversarial network (C-GAN) to generate synthetic tomato leaf images. Subsequently, the DenseNet121 model was fine-tuned through transfer learning on both synthetic and real image datasets to classify tomato leaf images into 10 disease categories.

Hassan et al. (2021) demonstrated improved computational efficiency by replacing standard convolutional layers with depthwise separable convolutions. The models were trained on an open dataset comprising 14 plant species and 38 disease categories, including healthy samples. Several hyperparameters—batch size, dropout rate, and number of epochs—were systematically optimized to evaluate model performance, yielding promising results.

Chen et al. (2021) employed MobileNetV2 pretrained on ImageNet and incorporated an attention mechanism to capture interchannel dependencies and spatial feature salience. They further optimized the loss function and applied a two-stage transfer learning approach during training. This proposed method outperformed existing approaches, achieving an average identification accuracy of 99.67% on a publicly available plant disease dataset. Simhadri and Kondaveeti (2023) employed a transfer learning strategy using 15 pretrained CNN models for automatic identification of rice leaf diseases. Among the models evaluated, InceptionV3 achieved the highest performance. Nagaraju and Chawla (2023) proposed a novel deep CNN architecture, NPNet-19, explicitly designed for maize plant disease classification. The model was evaluated on an expanded dataset comprising 15,960 images across six disease categories and one healthy class, primarily collected from maize fields in Telangana. During training, NPNet-19 achieved 97.51% accuracy, while testing yielded 88.72% accuracy, reflecting consistent yet realistic generalization performance.

Dong et al. (2023) proposed a set of widely used pretrained models aimed at improving the diagnostic accuracy of plant disease image classification. The models were evaluated across multiple plant disease diagnosis tasks. Experimental results showed that the pretrained models outperformed existing architectures in terms of accuracy while requiring less training time. Assad et al. (2023) introduced AppleNet, a DL-based multiclass classification model for apple plant diseases. The model efficiently extracted features from a real-world dataset using transfer learning with the pretrained ResNet50 CNN on the ImageNet dataset, thereby saving computational resources and time. By fine-tuning hyperparameters and training on 2897 augmented images, AppleNet achieved a classification accuracy of 96.00% for apple disease identification. Pradhan et al. (2022) conducted a comparative study of 10 widely recognized CNN architectures to evaluate their effectiveness in apple disease classification. Results demonstrated that DL techniques achieved high classification accuracy, with DenseNet201 outperforming all other models by attaining 98.75% accuracy.

Ahad et al. (2023) examined the rice disease classification capabilities of six CNN-based architectures. Transfer learning was explicitly applied, and an ensemble model named DEX was developed, integrating DenseNet121, EfficientNetB7, and Xception architectures. The evaluation employed a dataset comprising nine common rice diseases prevalent in Bangladesh. Sood and Singh (2023) applied transfer learning to pretrained CNN models—VGG16, ResNet50, InceptionV3, and DenseNet121—for plant disease classification using the PlantVillage dataset. Their findings underscored the effectiveness of transfer learning in leveraging existing pretrained models to enhance classification performance in agricultural applications.

Barbedo (2018) examined how dataset size and diversity influence the effectiveness of DL techniques in plant pathology. The investigation used an image database containing 12 plant species, each with varying sample counts, disease frequencies, and environmental conditions. Nader et al. (2022) compared the performance of several pretrained CNN models using transfer learning and ensemble learning strategies. Their methodology employed ensemble learning, combining three widely recognized CNN architectures to enhance classification accuracy for grape disease detection.

Adnan et al. (2023) proposed a method employing pretrained deep CNNs to classify a diverse dataset comprising 52 categories of plant diseases and healthy leaf samples. The study evaluated models including Xception, ResNet50, InceptionResNetV2, and InceptionV3, integrated with EfficientNetB3 and an adaptive augmented DL strategy, achieving high accuracy across multiple disease categories. Hayit et al. (2023) investigated the use of pretrained CNN models for early detection of Fusarium wilt infection types in chickpeas. Using a novel dataset of infected chickpea plants, the DenseNet201 model achieved an average test accuracy of 90%.

Zia Ur Rehman et al. (2021) employed two pretrained DL models combined with image enhancement and augmentation techniques to expand the citrus disease dataset. Hybrid contrast stretching was applied to enhance visual quality, followed by transfer learning to fine-tune the models. Additionally, feature fusion and optimization using the whale optimization algorithm (WOA) were applied to select the most relevant features, achieving a classification accuracy of 95.7% across six citrus disease categories. Chen et al. (2020) likewise applied transfer learning in their study. They employed pretrained models—VGGNet and Inception—originally trained on the ImageNet dataset. Instead of random weight initialization, the researchers leveraged knowledge from these pretrained networks. Their model demonstrated significantly improved performance compared to other established techniques.

Arshaghi et al. (2023) employed CNN-based approaches to classify potato diseases. Using a dataset of 5000 potato images, the proposed DL method was evaluated against established architectures, including GoogLeNet, AlexNet, R-CNN, VGG, and transfer learning frameworks. Results indicated that the proposed method outperformed existing models, achieving classification accuracies of up to 100%.

Thangaraj et al. (2021) aimed to enhance tomato plant disease detection by investigating the impact of different optimizers within a transfer learning framework. The system was evaluated using both real-time and stored tomato leaf images. Performance evaluations using Adam, SGD, and RMSprop optimizers revealed that the transfer learning approach was the most effective for automated classification of tomato leaf diseases. Uguz and Uysal (2021) explored the application of transfer learning using VGG16, VGG19, and a custom CNN architecture. A primary objective of their study was to examine the impact of data augmentation on model performance. Results indicated that models trained with data augmentation achieved up to 95% accuracy, compared to 88% for models trained without augmentation.

Some studies have integrated ML or optimization algorithms with DL models. Abd Algani et al. (2023) developed a DL technique, ant colony optimization-convolutional neural network (ACO-CNN), for plant leaf disease classification. Using ant colony optimization (ACO), the system extracted color, texture, and leaf features from images via a CNN-based classifier. The proposed approach outperformed existing methods, demonstrating higher diagnostic accuracy across multiple evaluation metrics. Alsubai et al. (2023) presented a hybrid DL model incorporating improved salp swarm optimization (ISSO). The method classified grape leaf images into four categories, beginning with median filtering (MF) for noise removal during preprocessing. The framework employed a dilated residual network (DRN) for feature extraction and utilized the Adam optimizer. Additionally, a convolutional neural network-gated recurrent unit (CNN-GRU) hybrid model was used for disease classification.

Verma et al. (2023) proposed a metalearning framework for plant disease classification, recommending top-n models for unseen datasets based on benchmark evaluations. Framework performance was evaluated using rank-biased overlap (RBO) to compare predicted and actual rankings. Extensive experiments with different metaextractor and metalearner configurations revealed that a probe network trained for 10 epochs, using standard deviation as the metaextractor and support vector regression (SVR) as the metalearner, outperformed other configurations. Garg and Alam (2022) proposed a hybrid architecture combining a pretrained CNN with a long short-term memory (LSTM) network. Transfer learning was applied to extract deep features from the fully connected layers of pretrained models, including InceptionV3, VGG16, and Xception. These features were concatenated with LSTM outputs and passed through a fully connected layer to enhance the model’s attention to salient information. The integrated model was then applied to classify apple foliar diseases.

Chug et al. (2023) introduced an innovative hybrid framework integrating ML and DL, featuring 40 hybrid DL models. These models incorporate eight pretrained EfficientNet variants (B0–B7) as feature extractors, coupled with ML classifiers including RF, LR, KNN, AdaBoost, and stochastic gradient boosting. Tabbakh and Barpanda (2023) designed a hybrid model named TLMViT, which combines transfer learning and vision transformer (ViT) architectures for plant disease identification. The TLMViT framework includes data collection using the PlantVillage and wheat datasets, image augmentation to reduce overfitting, feature extraction based on pretrained and ViT models, and final classification through a multilayer perceptron (MLP) classifier. Altalak et al. (2022) proposed a tomato leaf disease classification model that combines a CNN with a CBAM and a SVM.

Bajait and Malarvizhi (2023) proposed a grape leaf disease classification framework consisting of several key stages. The process begins with image preprocessing, including contrast enhancement using contrast limited adaptive histogram equalization (CLAHE) and noise reduction with adaptive bilateral filtering (ABF). Feature extraction was performed using the SqueezeNet model, and hyperparameter optimization was conducted with the equilibrium optimizer (EO) algorithm. Classification was subsequently carried out using a stacked autoencoder (SAE) model. Simulation analysis of the EODTL-GLDC technique on newly introduced plant disease datasets demonstrated highly promising results.

Figure 5 illustrates the performance of DL models applied to plant disease classification in the absence of visualization-based techniques. These models rely exclusively on raw image data and transfer learning architectures, excluding interpretable components such as saliency maps or class activation overlays. Notably, several models, including DenseNet121, EfficientNetB0, and InceptionResNetV2, achieved over 99% classification accuracy, underscoring the robustness of pure CNN-based pipelines. This finding highlights that high-performance plant disease diagnosis is achievable even without visualization-enhanced interpretability methods.

Tables 2 and 3 summarize the reviewed studies, highlighting the key models, accuracy rates, and principal contributions.

Studies employing visualization techniques—such as saliency maps, heatmaps, and feature attention mechanisms—have shown notable improvements in both interpretability and accuracy. Brahimi et al. (2018) employed saliency maps with the CaffeNet architecture, achieving 99.76% accuracy, whereas Atila et al. (2021) leveraged EfficientNetB5/B4 with augmented datasets, achieving up to 99.97% accuracy.

Other studies, including Nagasubramanian et al. (2019) and Wiesner-Hanks et al. (2019), explored 3D-DCNN architectures and UAV-based CNN classification, demonstrating that incorporating heatmaps and hyperspectral imaging enhances disease localization. These visualization-based methods provide an additional layer of interpretability, making them valuable for automated diagnostics and human-in-the-loop decision-making.

Studies focusing on direct classification without visualization techniques have emphasized optimizing model architectures, feature selection strategies, and training procedures to enhance performance. Naik et al. (2022) compared 12 CNN models, with VGG19 achieving 83.54% accuracy in chili disease classification. Several studies applied transfer learning, fine-tuning pretrained models such as ResNet, EfficientNet, MobileNet, and DenseNet for plant disease classification. Ahad et al. (2023) proposed an ensemble approach combining DenseNet121, MobileNetV2, and ResNet152V for rice disease detection, whereas Hassan et al. (2021) integrated depthwise separable convolutions, achieving up to 99.56% accuracy. Dai et al. (2024) introduced multilevel feature extraction methods, such as DFN-PSAN, to emphasize key disease regions in plant images, thereby enhancing classification robustness.

Several hybrid approaches integrate CNNs with metalearning and optimization algorithms to improve model adaptability. Garg and Alam (2022) developed a CNN–LSTM hybrid model for apple disease classification, leveraging sequential feature dependencies. Likewise, Abd Algani et al. (2023) applied ant colony optimization (ACO) in conjunction with CNNs to enhance color and texture feature extraction for leaf disease detection. Other metalearning strategies, such as the TLMViT model proposed by Tabbakh and Barpanda (2023), fused vision transformers (ViTs) with CNNs, enhancing the connection between spatial and contextual disease representations.

Despite substantial progress in DL-based plant disease classification, several challenges persist. Studies focusing on direct classification frequently encounter dataset biases and domain shifts, wherein models trained on controlled datasets often generalize poorly under real-world field conditions. In contrast, visualization-based methods—while offering improved interpretability—introduce computational overhead, rendering them less suitable for real-time deployment in resource-constrained environments. Future research should explore hybrid architectures that integrate self-supervised learning, generative adversarial networks (GANs), and multimodal fusion approaches—such as combining RGB and hyperspectral modalities—to overcome these limitations. Moreover, the integration of explainable artificial intelligence (XAI) frameworks can enhance model transparency and foster greater trust in DL-driven agricultural diagnostics.

3.3. Disease detection

Object detection in plant disease identification extends beyond simple classification to the complex task of spatial localization. In this domain, DL-based detection methods have evolved into two primary architectural paradigms: two-stage and one-stage networks.

One-stage detectors—such as the single shot multibox detector (SSD) and you only look once (YOLO)—have gained prominence due to their balance between inference speed and detection accuracy. Recent studies (Yin et al., 2024; Ali et al., 2024) have demonstrated the effectiveness of YOLO variants in detecting plant diseases, achieving high precision rates. Furthermore, integrating transformer-based architectures into detection frameworks has yielded promising results across diverse agricultural applications.

3.3.1. One-stage detection networks

The removal of the traditional region proposal stage has led to a paradigm shift in the development of single-stage object detection algorithms. These methods streamline the classification and regression processes by integrating the detection head directly within the backbone network. This integration significantly enhances the inference speed of the detection network. Among single-stage detection frameworks, two prominent models—SSD and YOLO—are widely recognized. Both models share a common characteristic, utilizing the entire image as input to the network. Consequently, this approach demonstrates the capability of one-stage detection algorithms to perform complex tasks—such as object localization and classification—both rapidly and accurately.

Compared with conventional CNN architectures, the SSD employs VGG16 as its backbone and integrates a feature pyramid network (FPN) to extract multiscale features, thereby improving prediction accuracy. Qiang et al. (2023) developed a dual-backbone enhanced single shot multibox detector (SSD) model for identifying citrus diseases. The detection accuracy and recall of the model were assessed, and its robustness was verified through detailed analysis of the results. The numerical results indicated that the trained network achieved an mAP of 72.54% on the test dataset. Furthermore, the model attained an mAP of 86.01%. Guo et al. (2021) proposed a precise identification method for common tomato diseases using DL. Their approach features a multi-resolution detector enhanced with optimized bounding box generation and assignment to improve feature extraction. The inclusion of dropout and the ADAMW optimizer effectively reduces overfitting. The detector was trained on images of healthy and diseased tomatoes and reliably detects 10 different diseases.

Saleem et al. (2022) developed a transfer learning-based approach to enhance plant disease detection through optimized network weights. The study compared multiple DL architectures to evaluate performance improvements. Weight optimization yielded a 9.25% improvement in mean average precision (mAP), achieving an overall mAP of 91.33%. In an earlier study, Saleem et al. (2020) evaluated three DL metaarchitectures—SSD, Faster R-CNN, and R-FCN—within the TensorFlow object detection framework for plant leaf disease recognition. Among these, the SSD model trained with the Adam optimizer achieved the highest mAP of 73.07%, demonstrating superior disease detection capability.

Sun et al. (2020) proposed an enhanced SSD framework for detecting corn leaf spot disease under complex environmental conditions. Their method emphasizes multiscale feature fusion through CNNs. The pipeline comprises several stages, including data preprocessing, feature fusion, feature sharing, and final disease detection. The enhanced model substantially outperformed the original SSD, increasing mean average precision (mAP) from 71.80% to 91.83%.

Similarly, Bao et al. (2023) developed the adaptive spatial feature fusion network (ASFFNet) for detecting wheat scab. ASFFNet incorporates a feature enhancement module that combines global and local representations, thereby strengthening the network’s representational capacity. An adaptive feature fusion module subsequently merges these enhanced features across multiple scales, effectively addressing detection challenges arising from small disease regions and improving detection accuracy. Comparative evaluations demonstrated that ASFFNet outperformed state-of-the-art object detection algorithms—including SSD, RetinaNet, YOLOv3, and YOLOv4—in terms of average precision (AP).

In contrast, YOLO reconceptualizes object detection as a single-stage regression task, enabling end-to-end prediction through a unified CNN architecture. Leveraging global image context, YOLO directly predicts bounding boxes and class labels, substantially increasing detection speed while maintaining high accuracy. A summary of these studies is presented in Table 4.

3.3.2. Two-stage detection networks

Two-stage detection networks have emerged as a principle approach in plant disease identification, effectively integrating region proposal mechanisms with classification tasks to enhance detection accuracy. The fundamental distinction between two-stage and one-stage networks lies in their processing pipelines: two-stage models first generate candidate regions potentially containing lesions and subsequently perform object detection within those regions, whereas one-stage models directly predict object locations and classes without a separate proposal stage (Wang et al., 2025).

Faster R-CNN, one of the most widely adopted two-stage object detection frameworks, has demonstrated high effectiveness in identifying plant diseases and pests. The network extracts a feature map from the input image using a convolutional backbone. Subsequently, a region proposal network (RPN) computes anchor box confidence scores to generate candidate regions, referred to as proposals. Following ROI pooling, the feature maps corresponding to these proposals are forwarded through subsequent network layers to refine detection, achieving accurate lesion localization and classification (Shao et al., 2019).

Building on this framework, Bari et al. (2021) employed Faster R-CNN for real-time detection of rice leaf diseases. Their model incorporated an improved RPN architecture capable of accurately localizing objects and generating robust candidate regions. The model was trained using both publicly available datasets and proprietary field images to enhance robustness. The study focused on three major rice leaf diseases: rice blast, brown spot, and hispa. The proposed approach achieved detection accuracies of 98.09% for rice blast, 98.85% for brown spot, and 99.17% for hispa, demonstrating high reliability across disease categories.

Gong and Zhang (2023) applied the Faster R-CNN framework to detect apple leaf diseases and compared its performance with that of YOLOv3. Their findings indicated that Faster R-CNN outperformed YOLOv3 in detection accuracy, primarily owing to its region proposal mechanism, which enables precise localization critical for capturing complex leaf disease patterns. Alruwaili et al. (2022) introduced a modified Faster R-CNN variant, termed RTFRCNN, designed for real-time detection of tomato leaf diseases. The model efficiently processed both static images and video streams, achieving an accuracy of 97.42%, surpassing AlexNet (96.32%) and conventional CNN models (92.21%) while requiring less computational power.

Similarly, Ozguven and Adem (2019) employed the Faster R-CNN framework to detect sugar beet leaf spot disease. Their model achieved a detection accuracy of 95.48%, effectively handling the natural variation in leaf morphology and confirming the suitability of DL-based approaches for conventional crops such as sugar beet. Mu et al. (2022) proposed a Faster R-CNN model with a feature pyramid network (FPN) backbone for detecting weed seedlings in agricultural fields. The model achieved an accuracy exceeding 95%, demonstrating the adaptability of the Faster R-CNN architecture for both disease detection and weed–crop discrimination under complex field conditions.

A summary of these studies is provided in Table 5.

Figure 6 presents an overview of deep learning-based models for plant disease detection, systematically categorized into one-stage and two-stage architectures.

3.4. Image segmentation

Accurate detection of diseased regions on plant leaves is essential for reliable diagnosis and severity assessment. Object detection methods such as Faster R-CNN provide bounding boxes around affected areas but often lack the pixel-level precision required for detailed pathological analysis. Consequently, image segmentation techniques have gained increasing importance in plant disease analysis, as they enable precise isolation of infected leaf regions from healthy tissue.

In recent years, DL-based segmentation architectures such as U-Net, Mask R-CNN, fully convolutional networks (FCNs), SegNet, and DeepLabv3+ have been extensively adopted in agricultural research. These architectures offer the advantage of producing pixel-level segmentation masks that delineate infected regions, thereby supporting tasks such as monitoring disease progression, quantifying severity, and enabling precise treatment application. Furthermore, integrating segmentation with classification frameworks enhances the robustness and reliability of automated plant disease detection in real-world agricultural settings.

Mzoughi and Yahiaoui (2023) addressed the challenges of plant disease recognition using image-based DL approaches. Unlike conventional approaches that jointly classify disease–species pairs, they proposed a novel framework capable of identifying plant diseases independently of the host leaf species. This design enables the recognition of previously unseen plant species exhibiting known disease patterns. Moreover, instead of relying on entire leaf images, their method focuses on local symptomatic features to reduce contextual bias and enhance model generalization. The authors developed a hybrid system that integrates DL-based semantic segmentation and classification networks to extract infected regions and accurately identify corresponding diseases. Extensive experiments conducted on the PlantVillage dataset validated the effectiveness of utilizing local disease symptoms for classification. The proposed approach also achieved notable improvements on more complex datasets, such as IPM and BING, which include leaves captured under uncontrolled environmental conditions.

Xia et al. (2021) investigated DL–based image segmentation methods combined with unmanned aerial vehicle (UAV) imagery to detect pine wilt disease (PWD), a major ecological threat to pine forests. Using fixed-wing UAVs, they collected aerial imagery over pine forests in Laoshan, Qingdao, China, and complemented these data with ground-based surveys to obtain additional contextual information. A dataset comprising 2352 annotated samples of infected pine trees under varying background conditions was assembled. The study evaluated the performance of several state-of-the-art semantic segmentation architectures, including fully convolutional networks (FCN), DeepLabv3+, and PSPNet. Among the tested loss functions, focal loss outperformed Dice loss, improving the average intersection over union (IoU) from 0.656 to 0.701. DeepLabv3+ achieved the best segmentation performance, with an IoU of 0.720 and an F1-score of 0.832. The superior performance of the model was attributed to the atrous spatial pyramid pooling module and the encoder–decoder structure, which effectively captured multiscale contextual and spatial information. Notably, increasing the depth of backbone networks did not enhance segmentation performance, as neither ResNet34 nor ResNet50 proved optimal across models.

In another study, Zhang and Zhang (2023) proposed an improved U-Net architecture, termed MU-Net, to address the challenges of segmenting diseased leaf images characterized by irregular shapes, varying sizes, complex textures, and noisy backgrounds. To enhance segmentation performance, MU-Net incorporates residual blocks (ResBlocks) to mitigate gradient vanishing and explosion issues, while replacing conventional skip connections with residual paths (ResPaths) to improve feature transformation between the encoder and decoder branches. The integration of ResBlocks and ResPaths increases the network’s depth and representational capacity. Experimental evaluations on a dataset of diseased leaf images demonstrated that MU-Net achieved superior segmentation accuracy and computational efficiency compared to conventional approaches.

Wang et al. (2023) introduced a novel segmentation architecture termed MFBP-UNet. The model incorporates a multiscale feature extraction (MFE) module to enrich both detailed and semantic feature representations, along with a BATok-MLP component that employs tokenized multilayer perceptrons and dynamic sparse attention to effectively balance local and global feature extraction. In addition, a diffusion-based data augmentation strategy was employed to enhance the model’s robustness and training stability. Experimental evaluations demonstrated that MFBP-UNet outperformed existing segmentation models, achieving substantial improvements over the baseline U-Net across multiple performance metrics: mean intersection over union (mIoU) = 86.15%, mean precision (mP) = 93.53%, mean pixel accuracy (mPA) = 90.89%, and Dice coefficient = 0.922.

Similarly, Kaur et al. (2024) developed a hybrid DL model, termed Hybrid-DSCNN, to segment and detect tomato leaf diseases using the PlantVillage dataset. The dataset was systematically annotated, enhanced, and augmented to improve training efficiency. The Hybrid-DSCNN integrates pretrained U-Net and SegNet architectures with instance segmentation mechanisms to enhance object detection performance. The model enables semantic segmentation of both single and multiple disease types, allowing precise identification and classification. Comparative evaluations against modified U-Net, M-SegNet, and U-SegNet models demonstrated that Hybrid-DSCNN achieved superior performance across multiple evaluation metrics, including accuracy, precision, recall, intersection over union (IoU), and mean IoU (mIoU). A total of 1004 images were processed, yielding an accuracy of 98.24%, surpassing the performance of all competing models.

3.5. Real-world deployment and Edge AI

Although DL models have demonstrated high performance in experimental evaluations of plant disease classification, their deployment in real-world agricultural environments faces practical challenges related to computational resources, inference speed, connectivity, and power consumption. Edge computing and lightweight DL architectures have been proposed to address these challenges, enabling real-time, on-site diagnosis using resource-constrained devices such as mobile phones and embedded systems.

MobileNet, EfficientNet, and their respective variants have emerged as preferred architectures for deployment on such devices owing to their computational efficiency and compact design. Depthwise separable convolutions, as implemented in MobileNetV2, significantly reduce model complexity and computational cost while maintaining competitive accuracy. Chen et al. (2021) developed a MobileNetV2-based model enhanced with channel-wise attention and dual-stage transfer learning, achieving 99.67% accuracy on a public plant disease dataset and demonstrating excellent suitability for mobile deployment. Similarly, Hassan et al. (2021) demonstrated real-time inference capabilities using MobileNetV2 and EfficientNetB0, achieving up to 99.56% accuracy across 38 disease categories and 14 plant species.

Beyond accuracy optimization, model compression techniques—such as quantization, pruning, and knowledge distillation—have attracted significant attention for reducing computational complexity and memory requirements. Khan et al. (2023b) applied quantization to a MobileNetV3-Small architecture, reducing the parameter count to 0.93 million while maintaining a classification accuracy of 99.50%. The optimized model, exported in open neural network exchange (ONNX) format, supports deployment across diverse edge platforms, including mobile devices. Guan et al. (2023) proposed a custom lightweight architecture, Dise-Efficient, derived from EfficientNetV2, which achieved 99.80% accuracy on the PlantVillage dataset.

Karim et al. (2024) designed a modified MobileNetV3-Large model for real-time grape leaf disease detection and deployed it on an NVIDIA Jetson Nano edge computing device. The system achieved 99.66% training accuracy and 99.42% testing accuracy and incorporated gradient-weighted class activation mapping (Grad-CAM) for visual interpretability, thereby enhancing model transparency and practical usability under field conditions.

Data and datasets

4.1. Publicly available datasets

Training DL models requires large and diverse datasets, as both the quantity and variability of data significantly influence the effectiveness of the learning process. Robust datasets encompassing images from multiple plant species and disease categories enable DL models to accurately identify a wide range of leaf anomalies. Dataset diversity is essential for developing models that generalize effectively to real-world conditions and minimize overfitting to specific environmental or imaging conditions. This section provides an overview of the most widely used plant disease datasets reported in the literature. Table 6 summarizes publicly available repositories, while Figure 7 illustrates the distribution of commonly used datasets according to their image counts.

4.1.1. PlantVillage dataset

The PlantVillage dataset1, introduced within the domains of agriculture and computer vision, has gained widespread recognition. It has become a cornerstone for developing and training DL-based models targeting key tasks such as plant disease identification and severity estimation, comprising an extensive collection of 61,486 images.

This extensive repository distinguishes itself as one of the largest and most comprehensive datasets available in the field. The dataset includes 14 crop species and 39 distinct disease classes, effectively capturing the variability and complexity of plant health conditions.

The PlantVillage dataset is distinguished by its diverse sources, incorporating contributions from both researchers and farmers. This multisource approach ensures that the dataset reflects the real-world complexity of plant diseases by capturing variations in image quality, disease stages, and environmental conditions. Each image in the repository is meticulously labeled and annotated with identifiers specifying the corresponding plant disease. Accurate labeling is fundamental in supervised ML, providing the foundation for training DL models to identify and classify plant diseases with high precision.

4.1.2. Leaf Image database

The leaf image database2 consists of 4503 unique images that have been systematically curated and categorized for research purposes. Of these, 2278 images depict healthy leaves, whereas 2225 correspond to diseased samples. The dataset includes 22 distinct categories representing various plant species and their respective health conditions. All images were captured under controlled laboratory conditions at Shri Mata Vaishno Devi College, Katra, India. The dataset is organized into two primary categories: healthy and diseased. Each image was further grouped according to plant species, labeled sequentially from P0 to P11. Subsequently, the entire dataset was subdivided into 22 subject categories, numbered from 0000 to 0022. Categories 0000–0011 correspond to healthy leaves, while 0012–0022 represent diseased samples. This structured categorization provides researchers and practitioners with a comprehensive resource for analyzing leaf images across diverse species and health conditions.

4.1.3. LeafSnap dataset

The LeafSnap3 dataset consists of 30,866 images representing various leaf types. Of these, 23,147 are high-quality images captured under controlled conditions using both backlight and reflected-light techniques. The remaining 7719 images were collected in field conditions using mobile devices. These field images depict leaves in their natural, unpressed state. Each image in this dataset is meticulously labeled with the corresponding tree species to which the leaf belongs. The dataset encompasses 185 distinct leaf classes, providing extensive species-level representation. Additionally, the dataset provides segmentation masks for leaf images. However, the segmentation process occasionally encounters errors, producing fully black segmentation masks in certain instances.

4.1.4. ImageCLEF dataset

The ImageCLEF4 dataset focuses on 126 tree species native to the French Mediterranean region for plant identification research. It consists of 11,572 images categorized into three types: scans (57%), scan-like photographs (24%), and natural field images (19%). The training set contains 8422 images, including 4870 scans, 1819 scan-like photographs, and 1733 natural images, each provided with XML annotation files. The test set comprises 3150 images—1760 scans, 907 scan-like photographs, and 483 natural images—all supplied with XML annotation files.

4.1.5. PlantDoc dataset

The PlantDoc dataset (Singh et al., 2020) contains 2598 images representing 13 plant species and 17 distinct disease classes. Its development required approximately 300 human hours of meticulous annotation. To evaluate the dataset’s effectiveness, three DL models were trained for plant disease classification tasks. The results indicated that using this dataset improved classification accuracy by up to 31%.

4.1.6. Plant disease symptom image database

The image database of plant disease symptoms (PDDB) (Garcia Arnal Barbedo et al., 2018) contains 2326 images depicting 171 diseases and disorders across 21 plant species. Although the original dataset is sizable, it remains insufficient for training high-capacity DL models. To overcome this limitation, each image was subdivided according to defined criteria, expanding the dataset to 46,513 samples. Both the original PDDB and its expanded version (XDB) are publicly available for academic research.

4.1.7. Rice leaf disease dataset

The rice leaf disease dataset (Shah et al., 2019) serves as a valuable resource for researchers and practitioners specializing in rice pathology and agriculture. It contains 120 JPEG images of rice leaves affected by three major diseases: leaf spot, brown spot, and bacterial leaf blight. Each disease class comprises 40 images, facilitating the development and evaluation of ML models for disease classification and identification. Accurate diagnosis of rice leaf diseases is essential for effective management, helping farmers and researchers prevent outbreaks and maintain healthy crops. This dataset provides a foundation for developing automated disease detection systems aimed at improving rice productivity and plant health, making it a valuable asset for the research community.

4.1.8. Flavia leaf dataset

The Flavia dataset (Wu et al., 2007) is one of the most well-known benchmark datasets for leaf recognition. It contains 1907 leaf images representing 33 distinct plant species. Each image has a resolution of 1200 × 1600 pixels and is freely available for research use.

4.1.9. DiaMOS plant dataset

The DiaMOS plant dataset (Fenu and Malloci, 2021) is a comprehensive resource containing images captured throughout the full growing season of pear trees, from February to July. The dataset aims to provide representative samples covering the primary developmental and cultural stages of the plant. It contains 3505 images in total, including 499 fruit images and 3006 leaf images, making it suitable for ML and DL applications in classification and detection.

4.1.10. FieldPlant dataset

The FieldPlant dataset (Moupojou et al., 2023) consists of 8629 annotated field images of leaves collected from plantations in Cameroon. It focuses on detecting and identifying diseases affecting three major tropical crops: corn, cassava, and tomato. Notably, FieldPlant is the first publicly available dataset to include annotated cassava images for plant disease detection tasks. This dataset provides a valuable resource for training efficient DL models for plant disease detection under real-world field conditions, supporting the advancement of object detection approaches.

4.1.11. New plant disease dataset

The new plant disease dataset5 contains approximately 87,900 RGB images of healthy and diseased crop leaves, categorized into 38 distinct classes. The dataset is divided into training and validation subsets using an 80/20 split, while preserving the original directory structure. Additionally, a separate directory containing 33 test images is provided for prediction and evaluation purposes.

4.1.12. CCMT dataset

The CCMT dataset (Mensah et al., 2023) was specifically developed for detecting crop pests and diseases. It contains images collected from local farms in Ghana and is available in two formats: raw and augmented. The raw dataset comprises 24,881 images categorized as follows: 6549 cashew, 7508 cassava, 5389 maize, and 5435 tomato samples. The augmented dataset, divided into training and testing subsets, contains 102,976 images across 22 categories: 25,811 cashew, 26,330 cassava, 23,657 maize, and 27,178 tomato samples. All images were anonymized, validated by expert plant virologists, and made publicly available to the research community.

4.1.13. Plant pathology dataset

The plant pathology dataset (Thapa et al., 2020) contains 3645 images depicting various symptoms of apple foliar diseases. These images were collected during the 2019 growing season from commercially grown apple cultivars in an unsprayed orchard at Cornell AgriTech, Geneva, NY, USA. Among the 3645 RGB images, 1200 depict apple scab, 1399 show cedar apple rust, 187 display complex symptoms involving multiple diseases on the same leaf, and 865 correspond to healthy samples. The images were captured in field conditions using smartphones under varying illumination, viewing angles, surface textures, and noise levels. The dataset was manually annotated into four categories—cedar apple rust, apple scab, multiple diseases, and healthy leaves—with all annotations verified by an expert plant pathologist.

4.1.14. Citrus fruit and leaf dataset

The citrus dataset (Rauf et al., 2019) contains 759 images of healthy and diseased citrus fruits and leaves. The images were manually captured using a DSLR camera under the supervision of domain experts to ensure labeling accuracy. The diseased samples were categorized into five groups: black spot, canker, scab, greening, and melanose. All images were resized to 256 × 256 pixels with a resolution of 72 dpi. Citrus fruit images were captured directly from plants in their natural environment, whereas leaf images were acquired under controlled laboratory conditions against a uniform gray background.

4.1.15. BRACOL coffee leaf dataset

The BRACOL dataset (Esgario et al., 2019) comprises images of Arabica coffee leaves specifically developed for the identification and quantification of coffee diseases and pests. It includes 2147 images depicting leaves affected by biotic stresses such as leaf miner, coffee leaf rust, brown leaf spot, and Cercospora leaf spot. These images were collected throughout the year in Santa Maria, Marechal Floriano—a mountainous region in Espírito Santo, Brazil. Captured using five different smartphones, the images focus on the abaxial (lower) leaf surface, with samples placed on a white background under partially controlled lighting conditions. The data collection process was intentionally diversified to enhance dataset variability, and annotation was performed with the assistance of an expert in plant biotic stress identification.

4.1.16. Apple leaf disease detection dataset

The apple leaf disease detection (ALDD) dataset (Jiang et al., 2019) contains 2029 images of apple leaves affected by five distinct diseases. The images were collected under diverse weather conditions, introducing natural variability and complex backgrounds that may challenge certain experiments. Following annotation and data augmentation, the dataset was expanded to 26,377 images, improving its suitability for DL model training.

4.1.17. Plant disease recognition dataset

The open-access plant disease recognition dataset6 comprises 1530 images categorized into three classes: healthy, powdery, and rust. These categories represent different plant health conditions: healthy indicates disease-free leaves, whereas powdery and rust denote leaves affected by specific fungal infections. The dataset is divided into training, validation, and testing subsets, which are essential for developing and evaluating ML models. Automated recognition and diagnosis of plant diseases through image analysis are critical in agriculture and horticulture, enabling early intervention, crop preservation, and improved productivity. This dataset serves as a valuable benchmark for researchers and ML practitioners developing automated plant disease recognition systems.

4.1.18. Kashmiri apple disease dataset

The dataset was collected from orchards in the Kashmir Valley for educational and research purposes and is publicly available on the Kaggle platform. It comprises approximately 419 images depicting both healthy and diseased leaves. Data collection took place during May, June, and July, coinciding with the peak season of plant disease prevalence. All images were manually captured using digital cameras and mobile phones from various brands (Sharma et al., 2022).

4.1.19. Groundnut leaf disease dataset

The groundnut leaf disease dataset (Aishwarya and Reddy, 2023) consists of digital photographs of groundnut leaves collected in the Koppal region of Karnataka, India, with the assistance of a plant pathologist under natural field conditions. The images are categorized into six classes based on leaf condition: healthy, early leaf spot, late leaf spot, nutritional deficiency, rust, and early rust. After preprocessing, the dataset contains 10,361 images organized into six folders: healthy leaves (1871), early leaf spot (1731), late leaf spot (1896), nutritional deficiency (1665), rust (1724), and early rust (1474). This comprehensive dataset serves as a valuable benchmark for training and validating ML and DL models in groundnut leaf disease classification and recognition.

4.1.20. Arabica coffee leaf dataset

The Arabica coffee leaf datasets, JMuBEN and JMuBEN2, were collected under real-world field conditions at the Mutira coffee plantation in Kirinyaga County, Kenya, using a digital camera and guided by an expert plant pathologist. The JMuBEN dataset includes three compressed folders: one containing 7682 images of Cercospora, another with 8337 images of coffee rust, and a third with 6572 images of Phoma. Conversely, the JMuBEN2 dataset consists of two compressed folders: one with 16,979 images of leaf miner infestations and another with 18,985 images of healthy leaves. Combined, the two datasets contain 58,555 annotated images across five categories: Phoma, Cercospora, rust, healthy, and leaf miner. Together, these datasets provide a valuable benchmark for training and validating DL models designed to recognize and classify Arabica coffee leaf diseases (Jepkoech et al., 2021).

4.1.21. Tobacco disease dataset

The tobacco plant disease dataset (Lin et al., 2022) comprises 2721 images of tobacco leaves captured under real-world field conditions. The dataset serves a dual purpose, supporting both disease classification and leaf detection tasks. The disease classification subset provides a diverse collection of images, enabling researchers and ML practitioners to develop models capable of accurately identifying and classifying multiple tobacco leaf diseases. It also supports leaf detection, which is essential for tracking and monitoring plant health and growth in agricultural applications. The combination of these two tasks makes the dataset particularly valuable for addressing challenges in tobacco disease management and ensuring the health and quality of tobacco crops.

4.1.22. Robusta coffee leaf dataset

The RoCoLe dataset (Parraga-Alava et al., 2019) consists of 1560 images of Robusta coffee leaves captured under real-world field conditions at a single plantation using a smartphone camera. It includes images showing visible red mites and rust spots indicative of coffee leaf rust infection, as well as healthy leaves without visible symptoms. Each image is annotated with details regarding the leaf objects, their health status (healthy or diseased), and disease severity (the proportion of leaf area affected by spots). RoCoLe serves as a valuable benchmark for evaluating ML algorithms in image segmentation and classification tasks related to plant disease recognition.

4.1.23. Black gram leaf disease dataset

The black gram plant leaf disease (BPLD) dataset (Talasila et al., 2022) focuses on Vigna mungo (commonly known as Urad), one of the most important pulse crops cultivated in India. The crop is severely affected by diseases such as anthracnose, leaf crinkle, powdery mildew, and yellow mosaic, which cause substantial yield losses among farmers. To facilitate early detection and classification, a dataset of 1000 images was created, comprising five classes: four disease categories and one healthy class. The images were captured using cameras and mobile phones under natural cultivation conditions in Nagayalanka, Krishna District, Andhra Pradesh, India. Agricultural experts assisted in labeling and processing the images. This dataset serves as a valuable resource for researchers applying image processing, ML, and DL techniques for the automated diagnosis and classification of black gram leaf diseases, thereby supporting farmers in disease management and yield improvement.

4.1.24. Sunflower fruit and leaf dataset

This dataset consists of images of healthy and diseased sunflower leaves and flowers affected by downy mildew, gray mold, and leaf scar disorders. The images were manually captured between 25–29 November 2021, at the Bangladesh Agricultural Research Institute (BARI) demonstration farm in Gazipur, in collaboration with an agricultural domain expert. The data collection period coincided with the sunflower plants approaching full bloom—a stage when disease incidence was as its peak. The original dataset contains 467 images collected from the sunflower demonstration field. To increase dataset size and variability, data augmentation techniques were applied, expanding the total to 1668 images in the augmented version (Sara et al., 2022).

4.1.25. Bangladeshi crop disease dataset

The new Bangladeshi crop disease dataset7 encompasses a wide variety of plant species, with a primary focus on four major crops: corn, potato, rice, and wheat. It contains 14 classes and 13,024 images, making it a valuable benchmark for plant disease recognition and classification research.

4.1.26. Wheat disease dataset

This dataset comprises a modest yet representative subset of wheat disease images, totaling 999 samples. These images depict real-world wheat growth conditions and are categorized into five distinct classes: yellow rust, brown rust, septoria, mildew, and healthy leaves. It is important to note that this subset represents only a fraction of the full dataset, which contains 19,172 images across the same five classes. The smaller subset provides a practical starting point for researchers and ML practitioners to develop and test models for wheat disease recognition and classification. The full dataset, with its comprehensive diversity of images, serves as a more robust and extensive resource for addressing challenges in wheat disease detection and agricultural improvement (Long et al., 2023).

4.1.27. New Zealand fungal and plant disease collection

The New Zealand fungal and plant disease collection (PDD) (Wilton, 2019) represents a major biodiversity repository encompassing a wide range of organism groups. Within its extensive holdings, fungi constitute the most represented group, with 106,016 samples. In addition, the collection includes diverse groups such as protozoa (2343), chromista (1261), plantae (119), bacteria (113), and a small number classified as incertae sedis (11). The PDD’s extensive diversity makes it an essential resource for researchers studying and preserving New Zealand’s unique fungal and plant biodiversity, including potentially novel species and their ecological interactions.

4.1.28. Sugarcane leaf disease dataset

The manually collected sugarcane leaf disease dataset serves as an essential resource for researchers and agricultural practitioners. It contains 2569 images categorized into five distinct classes: healthy, mosaic, red rot, rust, and yellow disease. The images were captured using smartphones with varied configurations, ensuring diversity in the dataset. The variety of devices used for image capture reflects real-world field conditions, making this dataset robust for developing ML models and algorithms. Collected in Maharashtra, India, the dataset is region-specific, offering valuable insights into sugarcane diseases prevalent in that region. The balanced distribution of images across categories enhances its utility, making it an ideal resource for training and testing sugarcane disease detection models. With variations in image size and RGB format, the dataset realistically represents the challenges encountered in real-world applications, facilitating the development of accurate and adaptable DL-based disease detection solutions for sugarcane crops (Daphal and Koli, 2023).

4.1.29. Tomato leaf dataset

The tomato leaf image collection comprises two distinct datasets, each derived from existing image repositories. The first dataset contains tomato leaf images extracted from the PlantVillage database, comprising 10 categories—nine disease classes and one healthy class. This substantial dataset includes 14,531 images, each depicting a single tomato leaf with its natural background. To streamline and refine the dataset, certain categories were excluded, resulting in a more focused and balanced collection. All images were resized from 256 × 256 pixels to 227 × 227 pixels to optimize them for further analysis and model development. The dataset is organized for practical evaluation using a five-fold cross-validation approach, ensuring robust testing and validation of ML models on this resource. This dataset serves as a critical benchmark for developing and evaluating models aimed at detecting tomato leaf diseases and supporting related research efforts (Huang and Chang, 2020).

4.2. Data collection and acquisition

The performance and reliability of DL-based plant disease detection systems are highly dependent on the quality, diversity, and quantity of data used during both training and evaluation. Data collection represents the foundational step in developing robust DL models and includes the acquisition, annotation, and curation of datasets (Ferentinos, 2018). Figure 8 illustrates a conceptual overview of the data acquisition workflow for plant disease detection. The process begins with capturing images under both uncontrolled field environments and controlled laboratory conditions. These images are subsequently annotated using dedicated labeling tools such as CVAT, LabelImg, and Labelbox. Following annotation, the data undergoes a curation phase—comprising cleaning, classification, and expansion—to ensure model trainability and consistency. Finally, the curated dataset is used to train DL models for disease recognition and classification tasks.

Image acquisition for plant disease detection generally occurs in two primary contexts: under field conditions and in controlled laboratory environments. Field-acquired data exhibit natural variations due to factors such as inconsistent lighting, occlusions, complex backgrounds, and ambient environmental noise. In contrast, laboratory-acquired images benefit from uniform lighting and controlled backgrounds, resulting in cleaner datasets that may, however, lack real-world variability. Both approaches present distinct advantages and limitations, and an effective DL model must be capable of generalizing across these variations (Mohanty et al., 2016).

Another major challenge lies in the accurate annotation of disease symptoms, which frequently requires domain expertise due to the visual similarity among different disease manifestations. The annotation process may involve simple class labels, bounding boxes for object detection, or pixel-level segmentation masks for semantic segmentation tasks. Tools such as LabelImg (Mathew and Mahesh, 2022), CVAT (Yue et al., 2025), and Labelbox (Chaudhari et al., 2023) facilitate this process; however, the required human effort remains considerable.

Beyond annotation quality, dataset imbalance remains a persistent and critical issue. Certain disease classes are well-represented, whereas others suffer from limited sample availability, potentially distorting the model’s learning process. Although data augmentation techniques are widely employed to artificially enhance sample diversity, the collection of geographically diverse, real, and balanced datasets remains an essential priority for future research (Ojo and Zahid, 2023).

4.3. Data preprocessing pipeline

Data preprocessing constitutes a vital stage in the DL workflow for plant disease detection, directly influencing model accuracy, generalization capability, and robustness. Although raw images capture the real-world complexity of plant diseases, they frequently contain noise, inconsistencies, and irrelevant visual details that may hinder the learning process (Ojo and Zahid, 2023). Effective preprocessing transforms raw data into a standardized and model-compatible format, ensuring consistency while improving the signal-to-noise ratio. Figure 9 presents a structured and visually enhanced workflow outlining the data preprocessing pipeline for DL-based plant disease detection. The process begins with the collection of raw images from diverse sources, including field environments, laboratory setups, and unmanned aerial vehicle (UAV) systems. These images undergo several preprocessing transformations, including resizing to match model input dimensions, pixel normalization to stabilize training, and noise filtering to eliminate irrelevant visual artifacts. Subsequently, background segmentation is applied to isolate plant regions, often followed by conversion into perceptually enhanced color spaces such as hue–saturation–value (HSV) or lightness–A–B color spaces (LAB) to improve feature representation. The output of this preprocessing pipeline serves as the standardized input for DL models, exerting a significant influence on both classification accuracy and generalization performance.

Images collected from multiple acquisition devices—including smartphones, DSLR cameras, and UAVs—typically vary in resolution, format, and aspect ratio (Shafik et al., 2023). DL architectures, such as VGG16, ResNet, and MobileNet require fixed input dimensions; therefore, image resizing constitutes a necessary preprocessing step. Commonly used input sizes include 224 × 224, 227 × 227, and 256 × 256 pixels, providing a practical balance between spatial detail preservation and computational efficiency.

Normalizing pixel intensity values to standardized ranges, such as [0,1] or [−1,1], ensures numerical stability throughout the training process. This normalization process accelerates convergence by aligning the input data distribution with the initialization and learning rate configurations of neural networks (NNs) (Gursoy and Kaya, 2024). Field-acquired images often contain nonleaf elements—such as soil, sky, or surrounding vegetation—that act as irrelevant background clutter. Such background elements can misdirect the model’s attention and introduce undesired correlations during feature learning. To mitigate this issue, several studies have implemented denoising techniques such as bilateral filtering, Gaussian blurring, and median filtering, which effectively suppress image noise while preserving edge information (Bhujade et al., 2024).

Background segmentation—where the diseased leaf is isolated from the surrounding scene—has proven beneficial for improving detection and classification accuracy. Barbedo (2018) proposed a saliency-based segmentation approach, which significantly enhanced classification accuracy for tomato leaf disease detection. Other commonly employed segmentation methods include thresholding in the HSV color space, k-means clustering, and GrabCut, all of which have been shown to improve model performance in recognition and segmentation tasks. Color transformation represents another essential component of preprocessing in plant disease detection, as disease symptoms—such as yellowing, necrosis, and rust—manifest through distinct chromatic variations. Although the RGB color space is commonly used, it is not always the most optimal representation for capturing subtle disease symptoms. Converting images into alternative color spaces such as HSV, LAB, or YCbCr enhances model sensitivity to hue and saturation variations, thereby enabling more accurate color-based disease classification (Trivedi et al., 2022).

4.3.1. Data augmentation techniques

Data augmentation plays a vital role in training DL models for plant disease detection, particularly in mitigating class imbalance, reducing overfitting, and enhancing model generalization. Since many plant disease datasets contain a limited number of samples—especially for rare or region-specific diseases—augmentation techniques are essential for achieving robust and reliable model performance (Ngugi et al., 2024). Data augmentation artificially expands the diversity of the training set by applying geometric and photometric transformations to existing images. Common operations include horizontal and vertical flipping, rotation (e.g., 90°, 180°), random cropping, zooming, shearing, brightness and contrast adjustment, and color shifting. Figure 10 illustrates several augmentation techniques applied to a single diseased leaf image.

Numerous studies have employed these techniques to enhance the performance of DL-based plant disease detection models. Min et al. (2023) proposed a data augmentation strategy based on image-to-image translation with attention mechanisms, generating realistic synthetic images of diseased leaves that improved classification accuracy. Khare et al. (2024) introduced LeafNST, a neural style transfer (NST)-based augmentation framework that synthesizes object-level textures to address dataset imbalance in plant disease images. This method generated diverse and highly realistic synthetic images, significantly improving both model generalization and accuracy. Furthermore, Muhammad et al. (2023) explored diffusion-based augmentation and demonstrated its superiority over traditional GAN-based methods in classifying coffee leaf diseases. Collectively, these studies underscore the effectiveness of both conventional and generative augmentation techniques in improving the robustness and predictive accuracy of DL models for plant disease detection.

Evaluation metrics for deep learning models

Various evaluation metrics have been widely adopted in the literature to assess the performance of DL models applied to plant disease classification, detection, and segmentation tasks. This section outlines the most commonly used evaluation metrics suitable for assessing classification, detection, and segmentation performance.

5.1. Accuracy

Accuracy measures the overall correctness of model predictions and is defined as:

$$[eqn]$$

5.2. Precision, recall, and F1-score

Precision quantifies the proportion of correctly predicted positive cases among all instances predicted as positive:

$$[eqn]$$

Recall (also referred to as sensitivity) measures the model’s ability to correctly identify all actual positive instances:

$$[eqn]$$

The F1-score represents the harmonic mean of precision and recall, balancing the trade-off between the two metrics:

$$[eqn]$$

5.3. Specificity

Specificity reflects the model’s capacity to correctly identify negative cases and is defined as:

$$[eqn]$$

5.4. Confusion matrix

The confusion matrix provides a detailed representation of the model’s prediction outcomes for each class and is widely employed in evaluating multiclass classification performance.

5.5. ROC curve and AUC

The receiver operating characteristic (ROC) curve plots the true positive rate (TPR) against the false positive rate (FPR) across multiple threshold settings. The area under the curve (AUC) quantifies the model’s overall discriminative capability across all possible classification thresholds.

$$[eqn]$$

5.6. Intersection over union

The intersection over union (IoU) is a standard metric in object detection and image segmentation that quantifies the spatial overlap between predicted and ground-truth bounding boxes or segmented regions:

$$[eqn]$$

5.7. Mean average precision

The mean average precision (mAP) represents the mean of the average precision (AP) values computed across multiple classes and IoU thresholds, serving as a key evaluation metric in object detection tasks:

$$[eqn]$$

5.8. Rank biased overlap

In metalearning frameworks, rank biased overlap (RBO) measures the similarity between ranked lists of model recommendations, where values range from 0 (no overlap) to 1 (perfect agreement).

5.9. Training and inference efficiency

Several studies additionally report computational efficiency metrics—such as training time, inference latency, and model size—to evaluate the feasibility of real-time and edge-based applications.

Collectively, these metrics provide a comprehensive framework for evaluating model effectiveness, particularly in addressing real-world agricultural challenges such as class imbalance, multiclass classification, and spatial localization accuracy.

Discussion

Deep learning (DL) models employed for plant disease diagnosis encompass diverse architectures, learning paradigms, visualization strategies, and segmentation and detection techniques, alongside practical considerations for real-world deployment. The accumulated evidence highlights substantial advancements in this domain while simultaneously underscoring key trends, strengths, limitations, and potential directions for further research. Although the field of DL-based plant disease diagnosis has expanded considerably, it remains fragmented in terms of model selection, evaluation methodologies, and deployment feasibility.

Numerous high-performing architectures, such as DenseNet201, InceptionResNetV2, and ResNet-50, have demonstrated outstanding classification accuracy on benchmark datasets like PlantVillage. However, their substantial computational demands pose major challenges for real-time inference and mobile deployment. In contrast, lightweight architectures such as MobileNetV2 and EfficientNetB0 (Chen et al., 2021; Hassan et al., 2021) demonstrate that marginal sacrifices in accuracy can significantly enhance deployability on edge devices. Nevertheless, few studies have systematically compared these models under consistent experimental conditions, raising questions about the acceptable trade-off between accuracy and computational efficiency. For instance, the notable accuracy achieved by a modified MobileNetV3 model deployed on Jetson Nano (Karim et al., 2024) effectively illustrates this trade-off between accuracy and efficiency..

Segmentation is often regarded as an alternative to classification due to its capability to localize and quantify disease severity. However, not all applications demand pixel-level precision; for example, classification may suffice in commercial crop sorting, whereas segmentation becomes essential in precision spraying. Studies such as Mzoughi and Yahiaoui (2023) demonstrate the potential of segmentation; however, few explicitly quantify how this added granularity translates into tangible practical outcomes. Moreover, segmentation-based models often require pixel-wise annotations—a frequently overlooked bottleneck that significantly constrains scalability and large-scale deployment.

Visualization techniques such as gradient-weighted class activation mapping (Grad-CAM) and saliency maps have become standard tools for evaluating DL models (Karim et al., 2024). However, many implementations rely solely on visual inspection, without systematically assessing the impact of explainability on user trust, model debugging, or agronomic decision-making. Despite their appeal, current explainable AI (XAI) methods lack standardized interpretability metrics, making them largely qualitative rather than operationally actionable. Therefore, there is a pressing need for usability studies to assess whether farmers and domain experts perceive these visualizations as interpretable and practically valuable.

Generative adversarial network (GAN)-based image synthesis, as explored by Muhammad et al. (2023) and Abbas et al. (2021), is frequently proposed as a solution to the data scarcity problem in plant disease classification. However, few studies have rigorously evaluated whether GAN-generated images faithfully preserve the pathological features necessary for reliable classification. Without extensive testing on real-world and unseen samples, models trained with GAN-augmented data risk overfitting to artificial artifacts rather than genuine disease characteristics. Although GANs offer significant potential, they cannot replace the need for diverse, well-annotated datasets collected under real-world field conditions.

Recent studies have begun incorporating transformer-based attention mechanisms and vision transformer (ViT)-style architectures (Alirezazadeh et al., 2023; Tabbakh and Barpanda, 2023). However, transformer architectures are often data-hungry and tend to be overparameterized for relatively simple plant disease imagery. It remains unclear whether these complex architectures provide a tangible advantage over CNNs in plant disease detection or if their adoption merely reflects broader trends in computer vision research. Future research should evaluate transformer performance on multimodal or geographically diverse datasets, where their capacity for modeling complex spatial and contextual relationships may be better demonstrated. Field conditions introduce challenges such as noise, occlusion, uneven illumination, and device variability—factors that are largely absent in controlled laboratory datasets. Although several studies have simulated real-world deployment on devices such as Jetson Nano or smartphones (Karim et al., 2024), few have rigorously validated their models under diverse and uncontrolled field environments. Moreover, aspects such as inference latency, power consumption, and user interface design are rarely subjected to systematic evaluation. Without comprehensive empirical assessment, deployability remains an assumption rather than a demonstrated capability.

Chug et al. (2023) reported that images containing multiple or overlapping leaves limit the applicability and generalization capability of their model. The study also identified limited dataset size as a major constraint (Min et al., 2023), potentially affecting the model’s accuracy and robustness. Notably, the current research did not investigate the segmentation of individual leaves from clusters, a preprocessing step that could potentially enhance classification accuracy.

Another major challenge in plant disease detection is the high similarity of geometric and morphological features among different disease classes (Bari et al., 2021). Disease identification and classification are further complicated by low-contrast image information, chromatic similarities between healthy and infected regions, and the presence of image noise. Model performance may also degrade due to noise, blur, and other distortions present in input samples (Albattah et al., 2022).

Min et al. (2023) noted that the vanilla CycleGAN architecture exhibits limitations in accurately capturing the distinct textural characteristics of target diseases. They further expressed concern regarding shape preservation in input leaf images due to indiscriminate background transformations during synthesis. The authors also highlighted that collecting diverse disease samples may exacerbate class imbalance issues within the dataset.

Lee and Yun (2023) reported that model performance varies depending on the types of environmental variables considered and the duration of environmental monitoring, emphasizing the importance of parameter optimization. Developing a comprehensive and well-structured database is essential before these models can be effectively deployed across diverse agricultural scenarios. Model usability and predictive performance could be enhanced by integrating additional sensors to capture a broader spectrum of environmental variables, reflecting current limitations in data acquisition. The study did not account for nonenvironmental factors that may influence crop disease prediction, potentially limiting the model’s overall accuracy.

Despite achieving high accuracy, few studies have addressed model failure cases or uncertainty estimation. Real-world deployment requires systems capable of recognizing and communicating their own limitations. Additionally, there is limited exploration of critical topics such as domain adaptation, cross-crop transfer learning, and privacy-preserving model learning. Furthermore, research on continuous model adaptation in dynamic field conditions remains limited, even as disease distributions evolve due to climatic and geographical changes.

Conclusions and future work

This work provides a comprehensive synthesis of DL approaches for plant disease diagnosis. It encompasses key topics including classification, detection, segmentation, explainability, data augmentation, real-world deployment, and architectural innovations. The reviewed studies collectively demonstrate how DL has revolutionized plant disease recognition, while simultaneously revealing persistent limitations and open research challenges.

The application of DL methods to plant disease detection has gained substantial attention owing to their ability to automatically extract discriminative features from complex visual data, eliminating the need for manual feature engineering. The reviewed literature indicates that convolutional neural network (CNN)-based architectures have substantially advanced the field, achieving high accuracy in both classification and localization across diverse plant diseases and image datasets. Beyond standard classification tasks, segmentation frameworks have enabled precise identification of diseased regions at the pixel level. Simultaneously, object detection models offer real-time localization capabilities that directly support modern precision agriculture systems.

Moreover, the growing emphasis on interpretability through visualization techniques—such as Grad-CAM, saliency maps, and heatmaps—enhances user trust and model transparency. Hybrid and attention-based architectures further improve feature extraction by enabling networks to selectively focus on spatial and channel-specific disease cues—an ability particularly valuable when symptoms are visually subtle or context-dependent. Despite these impressive advances, bridging the gap between laboratory-level accuracy and reliable real-world performance remains a persistent challenge. Most models continue to be evaluated on controlled datasets with uniform lighting and minimal background variability—such as PlantVillage—which fail to fully represent complex field conditions. Environmental factors—such as occlusions, soil background noise, and varying illumination—significantly affect model performance when applied to real-world imagery.

The cost of annotating data for segmentation and recognition tasks remains prohibitively high, particularly for small-scale farmers and research institutions with limited access to large annotated datasets. The reviewed studies emphasize that without diverse, high-quality, field-captured image data, DL models tend to overfit to region-specific features, thereby reducing their generalizability across different crop types, geographical regions, and agricultural practices. Several research opportunities and practical challenges are expected to shape future exploration of DL in plant disease diagnosis. First, there is an urgent need to develop and openly share comprehensive, well-annotated, and diverse field datasets. Such datasets should include images captured under diverse environmental conditions, across multiple regions, and at different growth stages to enable models to generalize effectively beyond controlled laboratory settings.

Furthermore, model compression techniques—such as pruning, quantization, and knowledge distillation—will be crucial for developing lightweight and efficient DL architectures suitable for deployment on mobile and embedded devices in low-resource agricultural environments. Optimizing energy efficiency and inference latency is particularly critical for real-time diagnostic systems operating in regions with unreliable or limited internet connectivity.

Another promising direction involves integrating diverse and complementary data sources. Combining RGB imagery with hyperspectral, thermal, LiDAR, and UAV-based data can provide complementary spectral and spatial information, enabling models to better distinguish between biotic and abiotic stress factors and to detect early-stage symptoms imperceptible to the human eye. Advances in self-supervised learning hold great potential to reduce dependence on large labeled datasets by enabling models to learn robust feature representations directly from unlabeled data. Similarly, federated learning frameworks facilitate collaborative model training across multiple farms and research institutions without the need to centralize raw data. This approach preserves data privacy while promoting knowledge sharing across distributed agricultural networks.

Ultimately, the importance of explainability and transparency in agricultural AI systems cannot be overstated. Developing unified architectures that seamlessly integrate classification, detection, segmentation, and explainability will yield trustworthy and actionable AI systems for farmers, agronomists, and policymakers. Beyond algorithmic enhancement, interdisciplinary collaboration with plant pathologists and agronomists is essential to ensure that DL models remain both biologically meaningful and practically applicable across diverse agricultural environments.