Plant diseases destroy crops silently and relentlessly, often spreading across fields before any farmer notices the first telltale lesion on a leaf. A new comprehensive review and bibliometric analysis, published in Discover Artificial Intelligence, has now mapped an entire decade of research devoted to solving this problem with artificial intelligence. By sifting through 1,081 publications from 2014 to 2024, drawn from the Lens.org and Scopus databases, the researchers document how deep learning, and convolutional neural networks in particular, have moved from experimental curiosity to one of the most active frontiers in agricultural technology. The picture that emerges is striking: publication output in the field has climbed almost continuously, from a trickle of papers in the mid-2010s to more than 250 documents per year by 2022 and 2023, with partial 2025 data suggesting an even steeper acceleration toward nearly 400 documents.

The stakes could hardly be higher. Agriculture anchors the economies of many developing nations, and the rapid, often invisible spread of plant disease is among the most damaging challenges farmers face. Traditional detection relies on human inspection and laboratory testing, approaches the review characterizes as time-consuming, labor-intensive, error-prone, and poorly suited to large-scale farming. When symptoms are subtle, even trained experts can miss them, and by the time disease becomes obvious it may already have devastated yields and inflicted serious financial losses, with knock-on effects for food supply. Machine learning and deep learning offer a fundamentally different proposition: models that learn patterns from vast libraries of crop photographs, satellite imagery, drone footage, and sensor readings, then flag diseased plants earlier and more reliably than the human eye.

At the technical heart of this transformation sits the convolutional neural network, or CNN. Unlike classical machine learning approaches such as Decision Trees, Random Forests, Support Vector Machines, and K-Nearest Neighbors, which depend on hand-crafted features like color, shape, and texture, CNNs learn hierarchical feature representations directly from raw pixels. Early convolutional layers capture edges and simple textures; deeper layers assemble those primitives into complex patterns such as lesions, mildew patches, or rust pustules. The review highlights landmark architectures including AlexNet, VGGNet, ResNet, and InceptionNet, each engineered to capture fine image detail and classify diseases in plant leaves. These models, the surveyed literature shows, can detect even subtle symptoms that escape manual inspection, and studies consistently report strong performance across the standard metrics of accuracy, precision, recall, and F1-score.

One technique emerges as especially consequential: transfer learning. Rather than training a network from scratch, which demands enormous labeled datasets that agriculture rarely offers, researchers take models pre-trained on the massive ImageNet dataset and fine-tune them for plant disease tasks, freezing and adapting layers as needed. The bibliometric data underline its importance, with transfer learning appearing in 86 studies. Complementing it, data augmentation artificially expands limited datasets by generating image variations, while newer tools such as attention mechanisms, which direct a model’s focus to the most informative regions of an image, YOLO for fast real-time object detection, and U-Net for precise image segmentation, mark a clear shift in the 2022 to 2024 literature toward methods that improve both accuracy and real-time performance. Ensemble architectures combining pre-trained backbones such as InceptionResNetV2 and NASNetMobile have even been applied to cucumber leaf disease classification, and CNN pipelines are now being folded into broader precision-agriculture systems, including automated weed detection in sugar beet farming.

The breadth of crops covered is remarkable. Among the 48 representative papers summarized in the review, researchers describe deep learning systems for jasmine leaf spot, tomato disease using hybrid reinforcement and transfer learning, lightweight real-time detection for pigeon pea, blight in potato and tomato, sugarcane disease classification integrated with a mobile app, apple leaf disease, black gram classification, and grape and mango disease detection in precision-farming contexts. The review also traces applications to leaf blotch, powdery mildew, rust, and drought or nutrient stress. This ubiquity across crops and deployment scenarios, the authors suggest, is itself a driver of the field’s steady publication growth, as each new crop, region, and deployment constraint spawns fresh modeling challenges.

Methodologically, the study is notable for its transparency. The team combined broad keywords such as deep learning in plant disease detection with narrower technical phrases, using Boolean AND and OR logic to balance coverage and precision. Filtering proceeded in strict order: restricting to 2014 through 2024, limiting documents to journal articles, books, and book chapters, excluding non-English work, deduplicating across the two databases, and manually removing out-of-domain papers that matched keywords but not the subject. Conference papers were deliberately excluded because peer review in that venue is inconsistent. The result was a corpus of 1,081 papers, of which 1,074 were peer-reviewed journal articles, with just seven editorials, letters, and news items retained because they met the same search criteria and constituted less than one percent of the total.

The bibliometric analysis reveals a genuinely global research enterprise. China leads in document count, a dominance the authors link to national smart-agriculture policy and one of the world’s largest pools of AI and computer science researchers, though they caution that volume does not equal quality or citation impact. The United States ranks second and India third, with Australia, the United Kingdom, Germany, South Korea, and Canada also contributing substantially. Co-authorship network analysis using VOSviewer identified the most connected researchers, with Babar Shah and Farman Ali each publishing seven documents and achieving a total link strength of 34, followed by Tariq Hussain and Muhammad Shoaib. Alvaro Fuentes and Sook Yoon stand out for combining seven documents with the highest citation count of 1,256, despite working in smaller, more focused teams. Color-coded cluster maps show dense international collaboration, with researchers from multiple countries frequently publishing together.

Keyword co-occurrence analysis adds a surprising cross-domain insight. In the broader network generated directly from the database exports, deep learning appeared 687 times with a total link strength of 2,671, machine learning 293 times, and convolutional neural network 118 times. But medical terms also surfaced prominently, including COVID-19, medical imaging, breast cancer, brain tumor, and diabetic retinopathy, reflecting how the same CNN architectures that identify lesions on plant leaves also detect tumors in X-rays, MRI scans, and CT images. The authors are candid about this limitation: their network maps were built from unfiltered search exports rather than the strictly agricultural 1,081-paper corpus, so keyword and co-authorship patterns characterize the wider deep-learning literature. They flag rerunning the maps on purely agricultural data as future work. Even so, the overlap is instructive, demonstrating that techniques refined for one diagnostic domain transfer readily to another.

The review closes with recommendations that read as a roadmap for the next decade. Governments, universities, and research groups are urged to build and share high-quality labeled datasets, since deep learning models perform only as well as the data behind them. Training pipelines, from online courses to hands-on workshops, need expansion to grow the skilled workforce the field demands. Crucially, the authors stress affordability: many farmers and rural clinics cannot afford expensive systems, so research should prioritize lightweight, low-cost models that run on mobile phones and low-power devices. Ethical and privacy safeguards, proper pre-deployment testing, and financial support for energy-efficient green AI round out the agenda. Taken together, the analysis portrays a field that has matured rapidly, connects laboratories across continents, and stands poised to deliver practical, real-time plant health monitoring, from smartphone apps to drone and IoT networks, as a cornerstone of smart farming and global food security.

Subject of Research: Deep learning and convolutional neural network applications for automated plant disease detection in agriculture

Article Title: A comprehensive review and bibliometric analysis of deep learning applications in plant disease detection

Article References: Sharma, P., Agarwal, P., Bairwa, A. K., & Sharma, V. (2026). A comprehensive review and bibliometric analysis of deep learning applications in plant disease detection. Discover Artificial Intelligence, 6(1), Article 1459. https://doi.org/10.1007/s44163-026-02310-z

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02310-z

Keywords: deep learning, plant disease detection, convolutional neural networks, bibliometric analysis, smart farming, transfer learning, precision agriculture, crop health monitoring, computer vision, agricultural AI, food security, machine learning