Answers are grounded in this paper and cite their sources.
Grapevine red blotch virus (GRBV) is a serious economic threat to grape production, yet vineyard-wide screening for GRBV is not feasible with current detection techniques of visual scouting and polymerase chain reaction (PCR) testing. To address real-world agricultural disease surveillance with limited labeled data, CompactSpectralViT, a lightweight vision transformer architecture for GRBV detection from UAV-based hyperspectral imagery, was developed. The study evaluated 714 PCR-confirmed vines sampled in four vineyards of Napa Valley, California, during two seasons of growth (2021 and 2024), utilizing hyperspectral imagery with 29 bands (520–820 nm) at 2-cm spatial resolution. A three-stage preprocessing pipeline standardized variable-sized vine canopies to uniform 64 × 64 × 29 hyperspectral cubes through adaptive resizing, content-aware cropping, and dimension adjustment, followed by two-stage band-wise normalization. CompactSpectralViT achieved 75.3 ± 2.8% accuracy with 87.2 ± 7.5% recall through 10-fold cross-validation. In input-matched comparisons, CompactSpectralViT achieved higher accuracy than the Support Vector Machine (62.9%) and Random Forest (65.1%) using 10 selected bands, and than SpectralCNN2D (72.0%) and HybridCNNTransformer (72.3%) using all 29 bands. It required 16.6 million multiply–accumulate operations per inference, approximately 7.6-fold fewer than the two full-band deep-learning baselines. Model interpretability analysis using SHAP (SHapley Additive exPlanations), corroborated by permutation importance, identified biologically meaningful spectral regions, including green (520–570 nm) for chlorophyll stress, orange-red (610–620 nm) for anthocyanin accumulation, and NIR (780–820 nm) for structural changes. Notably, wavelength selection experiments showed that a subset of 10 SHAP-selected bands achieved accuracy comparable to the full 29-band model (75.6 ± 3.1% vs. 75.3 ± 2.8%). The consistent cross-validation performance of CompactSpectralViT on a dataset spanning multiple commercial vineyard sites and seasons, together with interpretable feature selection, supports further development of scalable, affordable vineyard health monitoring systems that connect research innovation with operational disease management in precision viticulture.
No takes yet. Share an insight, caveat, or question.
Sanaeifar et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: