The ongoing strengthening of the greenhouse effect raises the Earth’s temperature and alters precipitation patterns. These factors significantly influence the occurrence, distribution, and transmission of plant leaf diseases. Pathogens and pests reproduce more quickly when temperatures rise (Chinnu et al., 2024). It leads to the rapid spread of pests and illnesses. Changes in precipitation patterns can lead to root damage or drought, lowering plant disease resistance. Plants are essential for managing the Earth’s climate. Plant diseases can worsen climate change, reduce crop quality and output, cause food shortages, and ultimately jeopardize human existence. Plant diseases can cause direct harm to plant tissues in a variety of ways, compromising plant aesthetics and market value. When it is severe, it can kill a huge number of plants. Early artificial identification of plant diseases had numerous limitations. Farmers have limited expertise. They have difficulties assessing the severity of disease progress. This would result in ineffective governance and excessive pesticide spraying, which will squander resources and pollute the environment (Strandberg et al., 2024). As a result, promptly, conveniently, and simply recognizing plant leaf diseases and improving plant disease control are critical for ensuring crop quality and yield, solving food concerns, sustaining human living conditions, and safeguarding the environment. This is also a significant motivation for studying this subject.

According to relevant information, there are many kinds of plant diseases. Plant diseases cause direct damage to plant tissues through parasitism, absorption of nutrients, destruction of cell structure, such as leaf spots, rot, fruit deformation, fruit cracking, etc. These damages not only affect the beauty of plants but also may reduce their market value; more serious ones can cause a large number of plant deaths (Roeswitawati et al., 2024). Accurate detection of plant leaf diseases is an important step in plant disease management. Due to the limitation of early technology, the identification of plant diseases is mainly manual identification, but due to the diversity and similarity of diseases and the low level of knowledge of some farmers, they cannot correctly judge the development degree of plant diseases, resulting in failure to prescribe the right medicine, thus causing the treatment process to be difficult to effectively control the development of plant diseases. In addition, in the process of treatment, a large number of pesticides are sprayed, which not only causes waste of resources but also causes environmental pollution and other problems (Abaineh et al., 2024). As a result, speedy, convenient, and simple detection of plant leaf diseases is critical to plant leaf disease control, as well as the field’s future development direction (Kalimuthu et al., 2021).

Although deep learning has achieved some results in identifying plant leaf diseases, it still faces numerous obstacles. How to improve the model’s generalization capabilities in multiple contexts and illness kinds while also better adapting to practical application scenarios. In the detection of plant leaf diseases, deep learning suffers from imbalanced data and erroneous labeling. This will reduce the model’s reliability and stability. Deep learning models that balance accuracy and computational resource consumption will be able to recognize plant diseases more efficiently on resource-constrained devices.

The proposed work discusses plant leaf disease recognition strategies based on on deep learning algorithms. Although deep learning did not attract widespread attention at the time due to hardware constraints, its potential gradually emerged. Krizhevsky et al. (2012) successfully implemented the famous convolutional neural network AlexNet on a graphics processor. At the Stanford ImageNet Large-Scale Visual Recognition Challenge, AlexNet made a breakthrough in classification accuracy far beyond other early computer vision methods. This accomplishment not only demonstrates deep learning’s formidable potential but also encourages the growth of diverse convolutional neural networks (CNN), quickening the field’s progress. Deep learning is being utilized more and more in the field of plant leaf disease identification, and its significance is becoming more and more apparent.

Deep learning is a branch of machine learning (Ganaie et al., 2022) that is based on artificial neural networks and learns complex data features through multiple hidden layers. Deep learning is the core technology to realize artificial intelligence, especially in the fields of image processing, speech recognition, natural language processing, and other significant breakthroughs (Matteo and J. Brendan, 2024). Deep learning models can automatically extract useful features from the original dataset, greatly reducing the burden of artificial feature recognition. And a deep learning model trained on a huge number of data points can learn the inherent law of data, even if new data does not show high prediction ability. Deep learning models may learn high-level and abstract data features, providing powerful representation capabilities (Rose and Rui, 2024). The subject of plant disease recognition has grown significantly as a result of deep learning intervention, with a number of reliable research outcomes emerging. Plant disease detection based on deep learning models has also become a hot spot in this field. Our work reviews the research and progress in plant leaf disease recognition based on deep learning models, as well as the challenges and opportunities faced by the field in recent years.

The suggested study examines the research and development of deep learning in plant leaf disease recognition. We present a thorough assessment of deep learning applications in plant disease recognition, taking into account recent research achievements. At the same time, we did a thorough review of the obstacles encountered in the field of plant disease diagnosis and investigated potential solutions. This establishes a clear path and reference for our future study. By summarizing and analyzing existing research findings, it is possible to encourage the further development of plant leaf disease identification technology, provide assistance in solving practical problems in plant disease control, and steer the field toward a more efficient, accurate, and practical direction.

We investigated convolutional neural networks, YOLO, structural optimization, and performance improvement strategies for lightweight models in disease recognition applications. We investigated the practical efficiency of several algorithms in difficult situations. It addresses accuracy, real-time performance, and the simultaneous detection of various diseases.

We mainly used Web of Science, the institute of electrical and electronics engineers (IEEE) Xplore, ScienceDirect, SpringerLink, and Google Scholar academic databases, focusing on literature from 2018 to 2025. This literature systematically reviews and integrates the research progress of deep learning in plant leaf disease recognition. The keywords we use are plant leaf disease, crop disease, deep learning, convolutional neural network, object detection, recognition, detection, and classification. Our inclusion criteria include deep learning-based leaf disease detection research, academic papers that use public or self-built field datasets, non-deep learning methods, pure theoretical models without image data, and research on non-leaf diseases such as root diseases. We acquired 580 papers after initial title and abstract screening and retained 210 following full-text reading. We also augmented 35 relevant articles with reference tracing. We supplemented and included over 20 prominent open-source datasets by searching Kaggle, GitHub, and academic institution official websites at the same time. We ultimately reviewed and integrated over 140 pieces of literature on algorithm creation, lightweight models, and multi-scenario applications.

The rest of the article is structured as follows: Section 2 offers a thorough analysis of each dataset and a thorough introduction to open-source databases both domestically and abroad. Plant species, sample size, and disease kinds are all included. Section 3 examines the specific performance and impacts of various implementation frameworks and provides an overview of plant leaf disease recognition algorithms based on one-stage detectors, two-stage detectors, and anchor-free frameworks. Section 4 examines the performance of alternative models and lightweight models under deep learning techniques. Section 5 concludes with a thorough analysis and forecast.

Deep learning models, based on large data sets, can learn the inherent laws of data at a deeper level and have a higher accuracy rate for predicting new data that has not yet been seen (Mingle et al., 2024). Thus, datasets are the basis for building deep models, and more quantitative and high-quality datasets tend to build more successful deep learning models (Alsaghir et al., 2022). A high-quality deep learning model can more accurately and effectively identify the type of plant disease, disease degree, and other information. Relevant individuals can provide the appropriate medicine based on the results of the deep learning model, resulting in more scientific and effective governance outcomes. Table 1 contains quality plant disease picture databases.

Plant leaf disease detection is the process of recognizing and locating infected areas, as well as the precise location of plant leaf disease in a complicated natural environment (Abade et al., 2021). This technology is the basis for realizing accurate classification and identification of plant leaf diseases and evaluating the damage degree of diseases. It is also the key link for accurately locating disease areas and guiding plant protection equipment to implement precise spraying.

Early plant disease target detection algorithms usually use a sliding window strategy to generate candidate regions, then extract the features of these regions and classify them with classifiers to finally determine the target regions (Ben et al., 2024). Viola-Jones detection, histogram of oriented gradient detection, and deformable component modeling are examples of common techniques. The sliding window method traverses the image by setting different scales and aspect ratios. Although this method can ensure that no potential disease regions are missed, it will produce a large number of redundant candidate windows, resulting in a significant increase in computational effort. In addition, it takes a lot of time to traverse the whole image, which makes the detection efficiency low. On the other hand, the feature extraction of candidate regions depends on manual design, and the extracted features mainly focus on the underlying information, such as the color and shape of diseases, whichamakes the robustness of disease detection poor. The a daptive boosting (AdaBoost) and support vector machine (SVM) are usually used in classifiers, but these methods have the problems of slow recognition speed and low accuracy.

The technical ability to process and analyze disease images in order to detect and differentiate between various disease categories is known as plant disease recognition technology. This technology serves as a crucial foundation for the prompt prevention and efficient treatment of plant diseases.

In the early stages of plant disease recognition research, feature extraction and screening relied heavily on artificial experience and domain knowledge. The core of its technical performance lies in whether it has two characteristics: one is whether it can fully express the essential information of disease targets and maintain high discrimination among categories. The other is whether it can form effective collaboration with classification algorithms. Traditional methods mainly use three kinds of visual features-shape, color, and texture-to classify and model. Texture features use gray level co-occurrence matrix, fractal dimension, Gabor filter, and other tools. Fractal dimension in particular is a useful tool for measuring the subtle variations in disease spot surface roughness.

Research practice shows that fusing multi-dimensional features and constructing a classifier ensemble system is the main technical path to improve accuracy. However, this artificial feature engineering poses a dual challenge to the practitioner’s professionalism: it requires both a deep accumulation of image processing technology and a knowledge reserve of plant pathology. Multiple interference elements, such as plant development stage difference, disease polymorphism, and environmental light change, make dynamic optimum adjustment of artificially built characteristic systems challenging to achieve. Traditional methods’ recognition accuracy tends to decline dramatically in complex field circumstances, particularly when the symptoms display atypical symptoms or multiple disruptions. This constraint derives mostly from the difficulties of using artificial feature modeling to adequately address the complexity and diversity of illness representations in real-world circumstances.

Deep learning learns feature representations directly from original image pixels through adaptive general algorithms, fundamentally breaking through the limitations of traditional artificial feature engineering. A number of significant convolutional neural network architectures have been developed as a result of changes in computing paradigms. These models include ZFNet in 2013, VGG and GoogleNet in 2014, ResNet in 2015, and later versions of DenseNet (2017), MobileNet (2017), and EfficientNet (2019). Through cutting-edge techniques like residual connection, deep separable convolution, and composite scaling, these network designs keep developing iteratively, progressively creating a technical ecosystem that can adjust to the demands of many scenarios. In general, deep learning-based techniques for identifying plant diseases adhere to the fundamental flow depicted in Figure 5 .

When using deep learning for plant pathology research in various agricultural settings, it is vital to select technical solutions flexibly based on individual circumstances and needs, as well as optimize practical benefits by merging cutting-edge technology. Because of available computational resources and stringent recognition accuracy criteria, the two-stage detection model should be preferred for high-precision analysis scenarios in the laboratory. High recognition accuracy can be achieved on typical datasets by divorcing the architecture of the regional recommendation network from the classification branch. However, it must accept the trade-off between high model complexity and poor inference speed, making it more appropriate as a core tool for disease mechanism research and precise diagnosis. Real-time monitoring scenarios in the field require a compromise between detection speed and dynamic environmental adaptability. The majority of detectors are either one-stage (YOLOv8, SSD optimized version) or anchor-free (CenterNet). The former considerably improves inference efficiency by implementing an end-to-end prediction method. The latter employs keypoint estimate rather than anchor box generation, ensuring great robustness in occlusion and complex backdrops. They can all support the real-time processing requirements of field mobile devices. Deep separable convolution and channel pruning approaches can be utilized to compress model parameters to less than 5 million, adjusting to the computational limits of low-power devices.In addition, the weakly supervised localization recognition method achieves lesion localization through image-level labels, further reducing annotation costs and providing a feasible solution for model deployment in resource-limited areas.

Future technological integration can concentrate on three key areas: To begin, we introduce the Transformer self-attention mechanism, which dynamically captures the long-range dependencies of disease regions, increasing model detection stability in scenarios with overlapping leaves and uneven lighting. Second, provide semi-automatic annotation tools that prioritize annotating high-value samples using self-learning methodologies, lowering annotation costs by more than 60% while maintaining data quality. Third, using neural architecture search (NAS) technology, heterogeneous network architectures are automatically generated for difficult field situations such as multi-scale illness spots and background interference. A lightweight architecture that merges dilated convolution and attention modules via reinforcement learning search achieves both accuracy and efficiency optimization. The following methodologies can help to improve the application of deep learning models in agricultural contexts, driving the progress of plant pathology technology toward precision and universality.

The actual field setting presents numerous obstacles, including reciprocal blockage of leaves, changes in sunlight, and complicated backdrops such as dirt and weeds. The available approaches are generally weakly resilient. When there is severe occlusion or an exceedingly complex backdrop, the performance of both the first-stage and second-stage detectors may suffer greatly. The anchor-free box approach uses center points or keypoints to provide relatively consistent performance under occlusion. Because of the lack of exact spatial information learning, lightweight recognition networks and weakly supervised algorithms are especially susceptible to background interference. The adaptation to occlusion improves the model’s robustness in complicated, changing field situations.

In terms of annotation cost, fully supervised object identification algorithms rely on a huge amount of precisely annotated bounding boxes and category information, and the annotation cost is quite high, making the technology difficult to implement. Lightweight recognition networks just need image-level labels like “disease” or “health,” as well as disease categories, which considerably reduces the annotation burden and makes them more feasible for large-scale classification systems. In scenarios when annotation resources are severely restricted, weakly supervised localization recognition techniques offer an alternate solution by utilizing image-level labels to find lesions. However, their localization accuracy and robustness are typically lower than fully supervised systems.

At the model deployment level, although researchers have successfully deployed models on edge devices, there are significant differences in computing power, storage capacity, and power consumption among different edge devices. This poses a great challenge to the adaptation of the model. Some models that perform well on high-performance devices may not be able to run at all or run at extremely slow speeds on resource-limited devices, which cannot meet the needs of real-time detection. For example, deploying complex deep learning models on low-end agricultural sensor devices with extremely limited resources still faces technical challenges. The maintenance and update mechanism after model deployment is also incomplete. The agricultural production environment is constantly changing, and new diseases continue to emerge. The characteristics of the original disease may also change. If the model cannot be updated in a timely manner, its detection accuracy will gradually decrease. However, model updates often require the collection of large amounts of data and retraining of the model, which is costly, time-consuming, and difficult to implement quickly on agricultural sites.

Performance trade-off is an unavoidable issue for deep learning models in agricultural disease detection. Accuracy, speed, and computational resource consumption are the three key indicators for measuring model performance. In practical applications, there is often a mutually restrictive relationship between them. Some models adopt complex network structures and a large number of parameters to improve detection accuracy. Although this enables the model to achieve high accuracy on specific datasets, it also brings problems of excessive computational resource consumption and slow detection speed. In agricultural fields, especially in large-scale farmland, it is necessary to quickly detect diseases in a large number of crops. Slow models cannot meet the actual production pace, which may lead to delayed detection and treatment of diseases, resulting in serious economic losses. On the contrary, in order to pursue detection speed and reduce computational resource consumption, some lightweight models have been proposed. However, these models often have insufficient accuracy and may result in false positives or false negatives. For example, in complex backgrounds or situations where early symptoms of diseases are not obvious, lightweight models may not be able to accurately identify diseases, which can affect disease prevention and control. How to find the optimal balance between accuracy, speed, and computational resource consumption is one of the urgent problems that deep learning needs to solve in agricultural disease detection.

The transferability of models to different regions is another major challenge for the application of deep learning in agricultural disease detection. There are differences in climate, soil, and planting variety factors in different regions. These differences can lead to variations in the types, patterns, and symptoms of plant diseases. Currently, most deep learning is trained and validated on datasets specific to a particular region. When these models are applied to other regions, their performance often deteriorates significantly due to differences in data distribution. For example, when the rice disease detection model trained in the southern region is applied to the northern region, it may not accurately identify the unique rice diseases in the north. In addition, the level of agricultural management and planting patterns in different regions can also affect the applicability of the model. Some regions may adopt advanced agricultural technologies and management methods. The crop growth environment is relatively stable, and the occurrence of diseases is relatively regular. Other regions may have outdated agricultural technology, complex and variable planting environments, and unpredictable disease occurrence. This makes the application of the same model in different regions require a lot of adjustment and optimization according to the local specific conditions, increasing the difficulty and cost of model application.

In conclusion, while deep learning has promising applications in agricultural disease detection, there are still many issues with model deployment, performance trade-offs, and model transferability between areas. Only through ongoing research and innovation to address these critical concerns will deep learning technology actually play a vital role in agricultural production and global food security.

With the help of deep learning technology, plant disease target detection and recognition studies have become a popular topic in scientific research circles both at home and abroad. A substantial body of research has greatly increased the accuracy of plant disease identification and contributed significantly to technological progress in this field. However, in actual implementations, these deep learning-based solutions encounter numerous challenges, and there is still a long way to go before fully and efficiently solving the problem of plant disease detection and diagnosis.

Data acquisition challenges: creating high-quality data takes a significant amount of time, labor, and money. Images of diverse plant varieties, disease types, stages of plant diseases, and other conditions should be gathered during the data set construction phase. In the processing step that follows, individuals with specialized knowledge should correctly identify the status of the plant disease.

Data imbalance: There is less data on plant diseases that are difficult to collect and rare than on plant diseases that are easy to collect and common, resulting in insufficient learning of deep learning models and affecting plant disease recognition accuracy.

Data noise interference: During the acquisition process, the image may be affected by external factors such as uneven natural lighting, a complex plant background, a poor shooting angle, and shooting height. These factors will interfere with the extraction and learning of plant disease features.

Computational resource requirements: Deep learning models have complex structures, and training and prediction require powerful computational resources. This increases the research threshold and research cost of plant disease identification.

Model optimization is difficult: Deep learning models have overfitting and underfitting problems, and hyperparameter adjustment requires a lot of experiments and rich experience. Problems such as vanishing or exploding gradients may also occur during model training, affecting the convergence and performance of the model.

Poor model interpretability: Because deep learning models are a black box, it is difficult to intuitively understand the models’ base and procedure for judging plant illnesses, as well as to troubleshoot and optimize.

Environmental adaptability: the field environment is complex and changeable, and the conditions, such as light, temperature, and humidity, are greatly different. The model performs well in the laboratory environment, but its performance may decrease in actual field application.

High real-time needs: large-scale agricultural output, the necessity for real-time disease detection, and the model’s reasoning speed all place increased demands on the model; some complicated models struggle to satisfy real-time requirements.

High deployment and maintenance costs: To deploy the model to the actual production environment, it is necessary to consider the compatibility of hardware equipment, network communication, data security, and other issues. Later model updates and maintenance also require professional technicians and resources.

Plant disease identification enters a promising era of change. In the field of deep learning, new models and algorithms are constantly emerging, such as transformers, lightweight neural networks, etc. With its excellent feature extraction ability, the deep learning method can capture the subtle features of plant diseases, thus significantly improving the accuracy and reliability of recognition.

With the advancement of sensor technology, plant physiological and environmental factors can be obtained more conveniently and comprehensively in the process of data collection so as to obtain higher quality plant disease data sets. In addition, multi-source data fusion creates conditions for building more accurate disease identification models. Plant leaf disease identification, a crucial component of precision agriculture, has drawn attention from a variety of sources, been supported by national resources, and accelerated the field’s growth as a result of the acceleration of the agricultural modernization process and the growing demand for precision agriculture.

Deep learning-based plant leaf disease detection has also seen significant prospects as interdisciplinary collaboration has grown. Combining computer science, agronomy, biology, and other fields allows us to examine plant disease occurrence settings, disease development characteristics, transmission rules, and identification techniques from a variety of perspectives. Because computer science has strong algorithms and data processing capabilities, it offers effective technical support for disease detection. From the standpoint of plant development environment and cultivation management, agronomy offers valuable practical expertise and important hints. Biology reveals the pathogenic mechanism of pathogens and the defense mechanism of plants by examining biological disorders at the microscopic level. The technique for identifying plant diseases will be advanced from various perspectives thanks to this multidisciplinary study. We have good reason to think that accurate early warning, comprehensive real-time monitoring, and effective plant disease prevention and management will soon be achieved, safeguarding both sustainable agricultural development and global food security.