Plant diseases significantly affect agricultural productivity and food security worldwide. Recent advances in

deep learning have enabled automated plant disease detection systems with high classification accuracy. Among

these approaches, convolutional neural networks (CNNs), ensemble learning techniques, and explainable

artificial intelligence (XAI) methods have emerged as promising solutions. This review presents a

comprehensive analysis of recent developments in plant disease detection from 2022 to 2026, focusing on deep

Agriculture plays a fundamental role in ensuring global food security and supporting the livelihoods of billions

of people worldwide. However, crop productivity is continuously threatened by various plant diseases caused

by fungi, bacteria, viruses, and environmental stress factors. According to reports from international agricultural

organizations, plant diseases account for significant annual crop losses, resulting in substantial economic damage

and reduced food availability. Early and accurate identification of plant diseases is therefore essential for

minimizing crop damage, improving agricultural productivity, and promoting sustainable farming practices.

Traditionally, plant disease diagnosis has relied on visual inspection by agricultural experts and plant

significantly transformed agricultural monitoring systems. In particular, deep learning techniques have

demonstrated remarkable success in image-based plant disease detection. Convolutional Neural Networks

(CNNs) such as AlexNet, VGGNet, ResNet, DenseNet, EfficientNet, and MobileNet have shown excellent

capabilities in extracting complex visual features from leaf images and achieving high classification accuracy.

The availability of large-scale agricultural image datasets and advances in computational resources have further

accelerated research in this area.

Despite their success, single-model deep learning approaches often face challenges when deployed in real-world

agricultural environments. Variations in illumination, background complexity, leaf orientation, disease severity,

and environmental conditions can significantly affect model performance. To address these limitations,

researchers have increasingly adopted ensemble learning techniques that combine predictions from multiple

where decision-making directly influences crop management strategies, explainability is particularly important.

Explainable Artificial Intelligence (XAI) techniques such as Gradient-weighted Class Activation Mapping

(Grad-CAM), Grad-CAM++, Local Interpretable Model-Agnostic Explanations (LIME), and Shapley Additive

Explanations (SHAP) have emerged as effective tools for visualizing and interpreting model decisions. These

techniques help identify disease-affected regions in plant images, thereby increasing user trust and supporting

practical deployment in agricultural environments.

Over the past few years, significant research efforts have focused on developing advanced plant disease detection

systems using CNNs, Vision Transformers (ViTs), ensemble architectures, and explainable AI frameworks.

However, the existing literature remains fragmented, with studies employing different datasets, evaluation

metrics, model architectures, and experimental protocols. As a result, it is often difficult to comprehensively

gaps, challenges, and future directions are identified to support the development of more accurate, robust,

interpretable, and field-deployable plant disease detection systems.

The major contributions of this review are summarized as follows:

models.

detection.

This review adopts a systematic approach to analyze recent developments in intelligent and transparent plant

disease detection using deep learning, ensemble learning, and explainable artificial intelligence (XAI). The

latest advancements in agricultural image analysis.

These databases were selected because they contain a large collection of peer-reviewed journal articles and

conference papers related to artificial intelligence, computer vision, machine learning, and smart agriculture.

The search process focused on studies published between 2022 and 2026 to ensure that the review reflects the

most recent technological developments and research contributions.

To retrieve relevant publications, a combination of keywords and search phrases was utilized. The major search

operators such as AND and OR to refine the search results and improve the accuracy of literature retrieval. The

search process initially produced a large number of publications covering various aspects of intelligent plant

publications.

To maintain consistency and quality, specific inclusion criteria were applied during the selection process. Studies

published between 2022 and 2026, written in English, and appearing in peer-reviewed journals or conference

proceedings were included in the review. In addition, selected studies were required to provide experimental

results and report standard evaluation metrics such as accuracy, precision, recall, F1-score, or Area Under the

Curve (AUC). Research articles focusing on image-based disease diagnosis of crops and plants were given

priority.

Similarly, exclusion criteria were established to eliminate studies that did not align with the objectives of this

categorized into major research groups. The first category includes conventional and advanced convolutional

neural network architectures such as AlexNet, VGGNet, ResNet, DenseNet, EfficientNet, and MobileNet. The

second category focuses on transformer-based approaches, including Vision Transformers (ViTs), Swin

Transformers, and hybrid CNN-transformer models. The third category covers ensemble learning techniques

that combine multiple models to improve classification accuracy and robustness. The fourth category reviews

explainable artificial intelligence methods such as Grad-CAM, Grad-CAM++, LIME, and SHAP, which enhance

the interpretability and transparency of deep learning systems.

To facilitate meaningful comparisons among different studies, several performance evaluation metrics were

analyzed. Accuracy was considered the primary metric for assessing overall classification performance, while

precision, recall, and F1-score were examined to evaluate model reliability and class-specific prediction quality.

In addition, some studies reported computational complexity, inference time, and deployment feasibility on

these datasets are widely adopted for evaluating plant disease detection algorithms. Comparative analysis of

these datasets helps identify their strengths, limitations, and suitability for real-world agricultural applications.

Through this systematic review methodology, a comprehensive overview of recent advancements in plant

gaps, and emerging opportunities in the field. Furthermore, it provides a foundation for understanding how deep

learning, ensemble learning, and explainable artificial intelligence can be integrated to develop accurate, reliable,

and transparent systems for next-generation precision agriculture.

datasets differ in terms of image quantity, disease categories, environmental conditions, image quality, and

annotation standards. Consequently, selecting an appropriate dataset is crucial for developing robust and reliable

plant disease detection models.

Among the available datasets, PlantVillage remains the most widely used benchmark dataset for plant disease

classification research. The dataset contains more than 54,000 annotated images representing healthy and

diseased leaves from multiple crop species. One of its major advantages is the availability of high-quality images

captured under controlled laboratory conditions with uniform backgrounds. Due to its balanced class distribution

and large number of samples, PlantVillage has been extensively used to evaluate the performance of

convolutional neural networks, transfer learning models, and ensemble learning approaches. Many studies have

reported classification accuracies exceeding 98% when trained and tested on this dataset. However, the

controlled imaging conditions limit its ability to represent real-world agricultural environments.

Another important dataset is the Plant Pathology dataset, which was developed to support research on apple leaf

disease classification. This dataset contains thousands of images representing various disease categories,

including scab, rust, and multiple disease combinations. The images were collected under diverse environmental

conditions and provide realistic disease symptoms encountered in orchards. The dataset has gained significant

attention through international competitions and benchmarking challenges, encouraging the development of

advanced machine learning and computer vision techniques for disease diagnosis.

In recent years, large-scale datasets such as PlantCLEF have emerged to address the need for greater diversity

in plant image analysis. PlantCLEF contains a vast collection of plant images collected from different

geographical regions, environmental conditions, and plant species. The dataset supports large-scale classification

tasks and provides opportunities for evaluating advanced architectures such as Vision Transformers and

multimodal learning systems. Its extensive variability makes it suitable for studying the scalability and

Besides these benchmark datasets, several crop-specific datasets have been introduced for diseases affecting

studies challenging.

While models trained on controlled datasets frequently achieve very high classification accuracy, their

conditions, leaf orientation, shadows, background clutter, disease severity, and image quality can affect model

predictions. Consequently, there is an increasing need for diverse and representative datasets that accurately

reflect practical agricultural scenarios.

Table 1 presents a comparative summary of widely used benchmark datasets for plant disease detection. The

comparison highlights key characteristics including the number of images, disease classes, environmental

conditions, and major applications. Understanding the strengths and limitations of these datasets is essential for

such as edges, textures, and color variations. As the network goes deeper, it begins to identify more complex and

disease-specific patterns that are important for accurate classification.

retaining the most important information. This is followed by fully connected layers that generate class

The final predicted class is:

To determine optimal weights, the ensemble is trained to minimize the cross-entropy loss over a validation

where:

subject to:

Since the objective function is differentiable, weights can be optimized using gradient-based optimization

methods such as:

Weights are iteratively updated as:

To prevent dominance of a single model and encourage balanced contribution, a regularization term can be

added:

This paper presents a multi-model ensemble framework designed to improve both the accuracy and transparency

of plant disease detection using deep learning. Instead of relying on a single model, the approach combines

multiple pretrained convolutional neural networks to capture different types of features from leaf images. This

helps the system perform more reliably, especially under varying environmental conditions. By using a weighted

decision fusion strategy, where each model contributes based on its strength, the framework reduces individual

model bias and improves overall consistency in predictions.

Along with better performance, the proposed system also focuses on making the results easier to understand. An

explainability component based on Grad-CAM is integrated into the framework to visually highlight the areas

affected regions rather than irrelevant background details, which increases user confidence and makes the system