Digital tools transform rice blast disease management, review finds

Rice blast, the fungal disease long regarded as the single most destructive threat to the world’s most important staple crop, has become the focus of an unprecedented wave of digital innovation, and a new systematic review now offers the clearest picture yet of what that technology can, and cannot, do. The analysis, published in the journal Smart Agricultural Technology, synthesized 109 peer-reviewed studies published between January 2021 and May 2026, spanning artificial intelligence, drones, remote sensing, the Internet of Things, and sensor-based diagnostics. Its central finding is striking: while laboratory accuracy in detecting the disease has effectively reached a ceiling, the technologies that could carry these tools into the hands of smallholder farmers remain largely unbuilt.

The stakes are enormous. Rice is the primary food source for more than 3.5 billion people, and global production must rise by an estimated 25 to 30 percent by 2050 to feed a projected population of 9.7 billion. Rice blast, caused by the fungus Magnaporthe oryzae, routinely destroys 10 to 30 percent of harvests and can wipe out entire crops under epidemic conditions. The pathogen attacks every above-ground part of the plant, using specialized infection structures called appressoria that generate enormous turgor pressure to physically rupture the leaf cuticle. Conventional defenses, resistant varieties, fungicides such as strobilurins, triazoles, and tricyclazole, and agronomic management, are steadily being undermined: the fungus’s high genetic variability rapidly overcomes resistance genes, and fungicide overuse fuels both resistance evolution and food-safety concerns.

The new review, led by Sandip Garai of the Indian Council of Agricultural Research’s Institute of Bioresources and Sustainable Agriculture with colleagues including Suryakant Manik, Kanaka K.K., Sudhir Kumar, Vijai Pal Bhadana, and Nishit Malviya, followed the PRISMA 2020 reporting standard. The team searched PubMed, Scopus, Google Scholar, and Semantic Scholar in May 2026, retrieving 2,056 records. After a two-pass deduplication combining exact DOI matching with fuzzy title matching, followed by automated keyword screening and full-text assessment, 109 studies survived. Screening reliability was checked on a random 30-record subsample by a second author, yielding a Cohen’s kappa of 0.86, indicating almost perfect agreement.

Artificial intelligence dominates the field. Of the 109 studies, 86 involved AI or machine learning, and 62 used convolutional neural networks, the deep learning architectures that revolutionized computer vision. Transfer learning from ImageNet-pretrained weights is nearly universal, with ResNet, VGG, MobileNet, InceptionNet, EfficientNet, and the YOLO family of object detectors all represented. The single highest reported accuracy in the corpus is 99.75 percent, achieved by ResNet50 classifying blast-infected versus healthy leaves on a 2,000-image dataset. A hybrid architecture called ResViT-Rice, which couples a ResNet backbone with a Vision Transformer encoder and attention modules, reached 99.04 percent accuracy while providing built-in explainability through attention maps. Object-detection studies, which localize lesions as well as classify them, span the full YOLO lineage from version 5 through 11 and include specialized variants such as Pyramid-YOLOv8, which targets small early-stage lesions, and YOLO-RBSD, which detects M. oryzae spores in microscopic images at 96.1 percent mean average precision and 125 frames per second.

Yet the review’s performance analysis reveals a sobering pattern. When 64 verified accuracy values were grouped across six detection tasks, median accuracy for laboratory-validated image classification studies has remained essentially flat, around 96 percent, since 2021, despite a flood of new architectures. Field-deployed studies report a median of 94 percent, a modest gap in absolute terms but one that masks the real problem: only nine studies across the entire corpus were validated in open fields. The authors conclude that the bottleneck is no longer model capacity but evaluation context, lighting variability, complex backgrounds, pathogen diversity, and growth-stage effects that laboratory datasets fail to capture. Notably, the correlation between dataset size and reported accuracy was only weak to moderate (r = 0.43, not statistically significant), suggesting that dataset curation, class balance, and evaluation rigor may matter more than sheer volume of images.

Beyond imaging, remote sensing is emerging as the most agronomically consequential frontier. Nineteen studies used ground-based spectroradiometers, hyperspectral cameras, UAV sensors, Raman and near-infrared spectroscopy, and satellites including Sentinel-2. The landmark capability is pre-symptomatic detection: identifying infection before any visible lesion appears. One team used machine learning-based feature selection to identify diagnostic wavelengths in the red-edge region between 700 and 740 nanometers, detecting blast at the asymptomatic stage with 92.9 to 94.8 percent accuracy. Another group developed a dedicated rice blast vegetation index, validated in UAV field trials with 95.9 percent overall accuracy, substantially outperforming the standard NDVI. Chlorophyll fluorescence imaging, which detects photosynthetic disruption rather than reflectance changes, offers similar pre-symptomatic power at roughly a tenth of the hardware cost of hyperspectral systems, a critical consideration for resource-poor farmers.

Drones and the Internet of Things round out the field-scale picture. Eleven UAV studies covered both detection and precision spraying; one showed that adding 1 percent lecithin adjuvant to drone-applied fungicide cut off-target drift by up to 3.16 times while increasing deposition along the spray path by up to 155 percent. On the sensing side, twelve IoT studies included platforms such as RiceBioS, which runs machine learning models on edge devices in low-connectivity fields and reduced latency by 40 percent compared with cloud processing, and AgriTalk, a cloud-fog-edge architecture that halved latency for farmers. Perhaps the most conceptually novel contribution is an electronic nose: a metal-oxide sensor array that detects volatile organic compounds characteristic of M. oryzae infection days before any visual or spectral signal appears, the only approach in the entire corpus capable of continuous, non-imaging, pre-visual monitoring.

Against these advances, the review documents three glaring absences. A systematic audit confirmed that the terms blockchain, digital twin, and federated learning appeared in none of the 2,056 records retrieved, not merely none of the 109 included studies. Blockchain has matured in agricultural supply-chain traceability elsewhere but has never been applied to rice blast surveillance or disease reporting; federated learning, which would allow institutions across blast-endemic regions to train shared diagnostic models without surrendering data ownership, is similarly unexplored; and digital twins, virtual crop-system replicas fusing sensors, weather data, and epidemiological models for scenario planning, are entirely absent. The authors project research horizons of 2025 through 2028 for closing these gaps, noting that the underlying technologies are already mature in other domains.

The review also scrutinized the quality of the literature itself. Using a purpose-built scoring framework called AI-Agri-QS, which rates dataset adequacy, validation rigor, metric completeness, and novelty on a 0-to-8 scale, the team found a mean score of just 1.50, with 89 percent of studies classified as low quality. Four systematic weaknesses recur: homogeneous single-source datasets shot under controlled lighting; roughly 40 percent of studies reporting only top-1 accuracy while omitting precision, recall, and F1-score; geographic concentration in East and South Asia with near-total neglect of sub-Saharan Africa and Latin America; and a complete absence of multi-season validation despite the year-to-year variability of blast epidemics. Only seven studies, 6 percent, incorporated explainable AI methods such as Grad-CAM or SHAP, a notable shortfall given that agronomists need to understand why a model flags a field.

The bottom line, the authors argue, is a field that is technically advanced but deployment-limited. The research agenda they propose centers on multi-site, multi-season field trials; affordable pre-symptomatic detection for smallholders; blockchain-enabled disease traceability; federated learning for privacy-preserving distributed training; explainable AI for trust and regulatory acceptance; and digital twin frameworks for epidemic simulation. For the farmers who bear the heaviest blast burden, many of them managing less than a hectare without reliable connectivity, the difference between a 99 percent laboratory score and a tool that works in a muddy paddy at dawn remains the decisive measure. The next generation of rice blast technology, the review concludes, will be judged not on benchmarks but on fields.