Deep Learning for Transformer-Based Plant Disease Detection: A Bibliometric Analysis ^†

Abstract

Agriculture, food security, and economic stability are impacted by plant diseases, making their identification and diagnosis essential. This article illustrates the research trends in plant disease detection using transformers through a bibliometric analysis based on visualization. The publications used in this work were collected from Scopus and Web of Science databases. For visualization, programs such as Biblioshiny and VOSViewer 1.6.20 were used. The results demonstrate that China is the most productive country, accounting for 11 total publications and 126 citations. China Agricultural University was the most productive institute, with six publications, while the Frontiers in Plant Science journal was the most productive journal, with six publications and 102 citations. It also demonstrates that the most used research topics in this field are “deep learning”, “plant disease”, and “vision transformer”. This study provides insights into the application of transformers for plant disease detection, enabling researchers to better understand and explore this field.

1. Introduction

Early detection and accurate diagnosis of crop disease play a vital role in protecting crop growth and increasing production. Crop disease detection and management help to improve crop yield and quality, as well as to optimize the use of resources [1]. Within the domain of deep learning (DL), pest and plant disease detection are critical; deep learning techniques have replaced traditional methods of plant disease identification [2]. Recently, deep learning techniques have demonstrated outstanding results in various domains, significantly enhancing the accuracy of object detection. Currently, deep learning is utilized to identify leaf diseases in crops like potatoes, citrus, wheat, and maize. These diseases include bacterial, viral, and fungal infections [3]. Manually identifying plant diseases is tedious and can cause errors in large cultivation areas. This requires a lot of time and labor. Accurate manual disease detection requires a lot of resources and a high level of skill. Early identification of plant diseases can reduce crop loss and pathogen transmission.

More recently, the use of transformers for plant disease detection has started to increase [4,5,6]. Several studies have employed transformers for the classification of grape diseases [7,8,9]. Aboelenin et al. [10] introduced a hybrid framework that combines convolutional neural networks (CNNs) and vision transformers (ViT) for the classification of plant leaf diseases. The authors utilized pre-trained models, to extract leaf features, while the ViT was employed to extract the deep features of the leaves. Apple and corn datasets from the PlantVillage dataset were utilized. The proposed method can successfully detect and classify plant leaf diseases, surpassing diverse cutting-edge models. Karthik et al. [11] presented a novel method for crop disease categorization using a dual-track architecture that integrates the swin transformer to extract global features, and a Dual-Attention Multi-Scale Fusion Network (DAMFN) that captures local features at multiple scales using specialized convolutional blocks. The CCMT dataset was used, which contains 102,976 images of various crop diseases. The proposed model had an accuracy of 95.68%. Maqsood et al. [12] proposed a customized architecture that combines vision transformers (ViT) and graph neural networks to accurately detect wheat diseases. It reached an accuracy of 99.20%. Chakrabarty et al. [13] developed an advanced model based transformer (BEiT model) for the detection of rice leaf diseases. Two public datasets were used. The model attained an accuracy score of 98.1%. Chen et al. [14] presented a hybrid architecture that integrates CNNs and transformers for the detection of tomato leaf diseases. The model was evaluated using three datasets. The authors utilized CyTrGAN, a Cycle-Consistent GAN-based model, to address data scarcity and class imbalance. The findings demonstrated that the proposed model, compared to ResNet, MobileNet, Swin Transformer, and Vision Transformer, achieves the best accuracy with fewer parameters and lower computational cost. Hamdi and Hidayaturrahman [15] proposed the use of pre-trained ViT models for plant disease categorization. Both the PlantDoc and PlantVillage datasets, were merged to create a final dataset comprising 38 classes and 56,857 images, addressing class imbalance issues. The results demonstrate that ensembles of Vision Transformer (ViT) models outperform individual models in plant disease classification, achieving a high accuracy of 98.16%. Remya et al. [16] proposed combining vision transformers (ViTs) with acoustic sensors to detect whitefly infestations in cotton crops. The authors utilized a custom dataset comprising 32,000 images and acoustic recordings collected from agricultural fields. Furthermore, the AgriPK dataset was used for experimental validation. The study demonstrated outstanding results for detecting cotton pests using vision transformers (ViTs) and acoustic sensors. The model attained an accuracy of 99%. Liu et al. [17] presented a NanoSegmenter model for high-precision tomato disease detection. It achieved 98% precision, 97% recall, and 95% mean Intersection over Union (mIoU), outperforming existing models such as DeepLabv3+, UNet++, and UNet. Ait Nasser and Akhloufi [18] presented a novel hybrid architecture to detect and classify plant foliar diseases. The authors used two datasets, Plant Pathology FGVC versions 7 and 8, and obtained accuracies of 98.28% and 95.96%, respectively. Wang et al. [19] presented a novel method to continual learning for crop disease identification using vision transformer total variation (ViT-TV), which achieved an accuracy of 95.38%, surpassing CNN architectures. Megalingam et al. [20] developed a classification system using vision transformers for accurate detection of diseases in cowpea plants, achieving an accuracy score of 96%.

This paper presents a bibliometric analysis of transformer-based plant disease detection. The aim of this work is to explore research trends in this field. It provides insights for future researchers by identifying the most relevant keywords, and the most influential journals and publications. Additionally, it analyzes journal productivity and citation impact. The rest of the paper is structured in this order: Section 2 outlines the methodology, including the publication selection criteria and the analytical tools employed. Section 3 presents the results and discussion, providing statistical insights into the distribution of research by institution, country, journal, and keywords. Finally, Section 4 concludes the paper and suggests future research directions.

2. Methodology

The Scopus and Web of Science databases were used to conduct a literature review for this study in January 2025. Articles with the keywords “transformers”, “plant diseases”, and “detection” in the title, abstract, or author’s keywords were found. Articles in the English language and all open-access criteria were included in the review. The timeframe was from “2015 to January 2025,” and the document types included “Article”, “Book Chapter”, “Conference Paper”, and “Review”. The method of publication selection is illustrated in Figure 1. The results were chosen using both the “BibText” and “plain text” formats. Each article was collected with the following information: title, abstract, source title, author, keywords, country, and institution. To conduct the bibliometric analysis, we used two analysis tools: Biblioshiny and VOSViewer. Biblioshiny 4.3.2 is an open-source R package 2024.12.0 used for bibliometric analysis. It offers a user-friendly interface for analyzing scientific articles, visualizing research trends, and exploring collaboration networks. The main features include importing data from sources such as Scopus and Web of Science, descriptive analysis of publications and citations, network analysis of co-citations, co-authorship, keyword co-occurrence, collaboration analysis, and graphical visualization that provides interactive plots, graphs, and charts for analyzing bibliometric data [21]. VOSviewer is a software program used for building and displaying bibliometric networks. These networks can comprise researchers, journals or individual articles, etc. It also offers a text-mining feature that can be used to create and display co-occurrence networks of significant phrases utilized in the scientific literature [22].

3. Results and Discussion

This section is dedicated to discussing the results of transformers for plant disease detection. The Transformers and Vision Transformers are both neural network models primarily used for natural language processing (NLP). The Transformer’s main innovation is the use of self-attention mechanisms, which enable the model to weigh the importance of different parts of the input when making predictions. The Vision Transformer is an extension of the Transformer model that handles images as a sequence of patches, instead of treating them as a grid of pixels. The existing studies in this field demonstrated outstanding results, which can motivate researchers using transformers for plant disease detection.

3.1. Annual Production

Analyzing the evolution of scientific productions, we observe that the number of publications related to plant disease identification, as shown in Figure 2, started to increase significantly from the year 2022, with 3 publications. In 2023, the number of publications increased to 9 and reached a peak of 23 publications in 2024.

3.2. Affiliations Analysis

The affiliations and number of articles contributed by each institution are listed in

Table 1. Most relevant affiliations.

Affiliation	Articles
China Agricultural University	6
Murdoch University	4
Sejong University	3
Vellore Institute of Technology (VIT)	3
VIT Chennai	3
Brac University	2
Egyptian Knowledge Bank (EKB)	2
Igdir University	2
Jilin Agricultural University	2
National University of Science and Technology Politehnica	2

, providing insight into their research productivity in a particular field. China Agricultural University contributed 6 articles, followed by Murdoch University with 4 articles. Sejong University, VIT, and VIT Chennai each contributed 3 articles, while five others contributed 2 articles each.

3.3. Journals Analysis

Table 2. Top 10 relevant sources.

Sources	Articles	Number of Citations	Start Publication Year
Frontiers in Plant Science	6	102	2022
Sensors	3	52	2023
Agronomy	3	43	2022
Agriculture (Switzerland)	1	21	2022
Computers and Electronics in Agriculture	2	16	2024
IEEE Access	2	3	2024
Computers	1	3	2024
AI	1	2	2024
Artificial Intelligence Review	1	1	2024
Smart Agricultural Technology	2	1	2024

demonstrates the most productive journals organized by citation counts. The Frontiers in Plant Science journal has the most citations (102) among its six publications. Sensors ranks second with 52 citations, followed by Agronomy (43 citations), Agriculture (Switzerland) (21 citations), and Computers and Electronics in Agriculture (16 citations). Figure 3 illustrates the trend of source production over time in the top six journals. Three journals, including Frontiers in Plant Science, Agronomy, and Agriculture (Switzerland), an publishing articles on DL-based plant disease detection in 2022. However, we can conclude, based on the number of citations, that the Frontiers in Plant Science journal has a greater reach for DL-based plant disease detection publications compared to the Agronomy journal and the Journal of Agriculture (Switzerland), even though the number of publications in these journals is almost the same in this field.

3.4. Countries Production Analysis

Table 3. Countries’ scientific production.

Country	Total no. of Publications	Percentage of Publication	Total Citations
China	11	30.5	126
India	10	33.3	12
Australia	6	16.6	15
Saudi Arabia	5	13.8	21
South Korea	4	11.1	32
Malaysia	3	8.3	3
Pakistan	3	8.3	22
Spain	3	8.3	1
USA	3	8.3	17

shows the most productive countries. China has the most articles published on DL-based plant disease detection, with 11 publications (30.5%) and 126 citations. India ranks second with 10 (33.3%) articles and 12 citations, followed by Australia with 6 (16.6%) publications and 15 citations. Saudi Arabia ranks fourth with 5 (13.8%) publications and 21 citations. It is observed that countries like Australia, Saudi Arabia, South Africa, and Pakistan have a smaller number of publications compared to India. However, they have a more significant number of citations than India due to the collaboration or outreach of publications produced by Australia, Saudi Arabia, South Africa, and Pakistan. In summary, transformer-based plant disease detection has been dominated by China and India.

3.5. Top Documents by Citations

The impact of articles on the literature is evaluated by document citation analysis. A citation analysis was conducted to identify the most influential publications and summarize their content.

Table 4. Most-cited global documents.

Paper	DOI	TC	TC per Year
Kunduracioglu, I., 2024, [9] j plant dis prot	10.1007/s41348-024-00896-z	14	7.00
Liu, Y., 2023, [17] plants	10.3390/plants12132559	10	3.33
Zhang, Y., 2022, [23] front plant sci	10.3389/fpls.2022.875693	63	15.75
Jajja, A., 2022, [24] agric	10.3390/agriculture12101529	21	5.25
Barman, U., 2024, [25] agronomy	10.3390/agronomy14020327	17	8.50
Rezaei, M., 2024, [26] comput electron agr	10.1016/j.compag.2024.108812	15	7.50
Parez, S., 2023, [27] sensors-basel	10.3390/s23156949	25	8.33
Li, G., 2023, [28] front plant sci	10.3389/fpls.2023.1256773	24	8.00
Shaheed, K., 2023, [29] sensors	10.3390/s23239516	21	7.00
Wu, J., 2022, [30] agronomy	10.3390/agronomy12092023	19	4.75

presents the top ten most influential publications, ranked by total citations (TC) and total citations per year.

Zhang et al. [23] conducted a study on automatic plant disease detection using Generative Adversarial Networks (GANs) and transformer architectures to enhance model performance. The authors used anchor-free object detection methods and data augmentation techniques to improve the accuracy of disease identification from leaf images. An agricultural robot was built to apply the model to the real world. The PlantDoc dataset was used, which contains 13 plant species and 27 categories. The findings showed that the model carried out the best performance compared to traditional models. It reached a precision of 51.7%, and recall of 48.1% on the validation set. Jajja et al. [24] suggested a compact convolutional transformer (CCT)-based method to detect whitefly attacks on cotton crops. The authors created a new dataset named AgriPK, which was categorized into five classes: healthy, unhealthy, mild infection, severe infection, and nutritional deficiency. Another dataset, the cotton disease dataset, was used for comparison purposes to evaluate the performance and generalization. The findings demonstrated that the proposed approach attained an accuracy of 97.2%, outperforming all other deep learning methods, including MobileNetv2, ResNet152v2, VGG-16, SVM, and YOLOv5.

3.6. Keyword Co-Occurrence Analysis

Keyword co-occurrence analysis in bibliometric analysis helps identify key topics, trends, and relationships within a research field. VOSviewer was used to create a keyword co-occurrence network, as shown in Figure 4. The minimum number of keyword occurrences is set to 4. From 207 saved keywords, 12 met the criteria and were divided into 2 clusters (red and green).

4. Conclusions

This paper presents a bibliometric analysis of transformer-based plant disease detection research trends. The existing studies highlight the efficiency of transformers for plant disease detection, showing impressive results in terms of accuracy. This study analyzed 36 publications published between 2021 and January 2025 from Scopus and Web of Science. We presented scientific maps that illustrate annual publication numbers, countries, journals, research institutes, and author productivity for DL transformer-based plant disease detection. The bibliometric analysis demonstrates that the most productive countries are China, India, Australia, Saudi Arabia, and South Korea. The most used topics include deep learning, classification, plant disease, vision transformer, and plant disease detection. For research institutes, China Agricultural University published more articles than other institutes. Frontiers in Plant Science was the most relevant source, with a maximum of 102 citations. The analysis also includes the most cited documents. In future work, we will look to search additional databases for relevant recent papers.