AI Drone Model Detects Grapevine Red Blotch With 75% Accuracy
A compact artificial intelligence model analyzing drone-collected hyperspectral imagery can identify grapevine red blotch virus across commercial vineyards with 75 percent accuracy and 87 percent recall, a study published in Smart Agricultural Technology reported.
Key Clinical Takeaways:
- Researchers developed CompactSpectralViT, a lightweight vision transformer that processes canopy hyperspectral cubes using 7.6 times fewer operations than standard convolutional baselines.
- The diagnostic model achieved an 87.2 percent recall rate across 714 geolocated vines in Napa Valley, catching nearly nine in ten infected plants to prevent vector-driven spread.
- Spectral analysis revealed that infection alters reflectance sharply in the 540 to 580 nanometer green region and the 740 to 760 nanometer range, while standard vegetation indices like NDVI failed to separate infected from healthy vines.
Pathogen Impact on North American Viticulture
Grapevine red blotch virus functions as an economically damaging pathogen in North American viticulture by disrupting sugar transport within berries, delaying fruit ripening, and muting the characteristic color and profile of wine. Traditional scouting relies on slow visual inspections and expensive laboratory testing, creating a significant diagnostic bottleneck for growers managing large acreages. The virus spreads via the three-cornered alfalfa hopper, making rapid identification critical to preventing wider vineyard contamination.
Hyperspectral Data Collection Across Napa Valley
A research team led by Alireza Sanaeifar of California State University, alongside Eve Laroche-Pinel, virologist Marc Fuchs, and Luca Brillante, conducted field trials across four commercial Cabernet Sauvignon and Cabernet Franc vineyards in Napa Valley, California. Over two growing seasons, the investigators geolocated 714 vines using satellite navigation, collected petiole samples from each plant, and confirmed infection status through endpoint multiplex PCR testing. This process yielded a balanced dataset comprising 399 infected and 315 non-infected vines. Aerial data collection utilized a DJI Matrice 600 Pro drone equipped with a Senop HSC-2 snapshot hyperspectral camera featuring 29 spectral bands spanning 520 to 820 nanometers. Flights executed at 30 meters altitude and 5 meters per second generated imagery with a ground resolution of 2 by 2 centimeters per pixel.
Infected Vines Show Specific Narrow Band Reflectance Changes
Detailed spectral evaluation demonstrated that infected and healthy vines share nearly identical overall reflectance curves, with statistically significant divergences isolated to specific narrow bands. Healthy vines reflected more light in the green region between 540 and 580 nanometers, reflecting chlorophyll degradation and pigment alteration caused by the pathogen. Conversely, infected vines exhibited localized reflectance increases between 740 and 760 nanometers, indicating structural leaf changes. While the standard NDVI vegetation index proved ineffective at differentiating infection classes, the Green NDVI and the Anthocyanin Reflectance Index successfully discriminated the groups, aligning with the anthocyanin accumulation characteristic of red blotch disease.
CompactSpectralViT Processes Hyperspectral Data via Vision Transformer
To overcome the limitations of blunt conventional vegetation indices, the team engineered CompactSpectralViT, a lightweight patch-based vision transformer designed specifically for hyperspectral data. The system standardizes each vine canopy cube to 64 by 64 pixels across 29 bands using proportional resizing, adaptive center cropping, and final adjustment. The cube is subsequently divided into 64 patches measuring 8 by 8 pixels, each containing all 29 spectral values, before passing through a two-stage embedding that compresses every patch to 48 dimensions. Four transformer encoder blocks with three attention heads each allow the model to evaluate canopy relationships simultaneously. Evaluated via 10-fold stratified cross-validation, the model attained a mean accuracy of 75.3 percent, a precision of 74.5 percent, an F1-score of 79.8 percent, and a ROC-AUC of 0.744. The high recall rate of 87.2 percent minimizes missed infections that could otherwise fuel viral transmission.
Computational Efficiency on Graphics Hardware
Benchmarking tests executed on an NVIDIA RTX 3090 demonstrated that the compact architecture requires only 16.6 million multiply-accumulate operations per inference, reducing computational demand by roughly 7.6 times compared to standard convolutional baselines and hybrid comparator models. The network consumed 138 megabytes of GPU memory, outperforming alternative architectures that required 376 and 440 megabytes while securing higher marks in accuracy, recall, and ROC-AUC.
Disclaimer: The information provided in this article is for educational and scientific communication purposes only and does not constitute medical advice. Always consult with a qualified healthcare provider regarding any medical condition, diagnosis, or treatment plan.