Herbicide-resistant weeds already cost UK wheat producers £400 million a year across more than 520 resistant biotypes, and a new study from the Royal Agricultural University suggests the choice of AI architecture will determine whether precision spraying can dent that bill in dense, partially hidden canopies.

The study, published in Scientific Reports on 23 June 2026 by Euan Hall, Emmanuel Junior Zuza, Karen Rial Lovera and Chris McCarthy, tested four image-classification architectures on 970 field images of three wheat-field weed species collected at Coates Manor Farm, Cirencester, between April and July 2025. The vision transformer ViT-B16 reached 86% test accuracy when half of each image was artificially occluded — up from 80% with no occlusion — while the CNN baseline EfficientNet-B0 collapsed from 66% to 46% under the same treatment.

Weeds cause wheat yield losses of 23–50%, and wheat supplies one-fifth of global human calories. The researchers frame reliable classification under partial visibility as the gating factor for spot-spraying systems, where a false negative means an unsprayed weed and a false positive means herbicide wasted on crop.

What did the researchers actually test?

The dataset covered wild oat (Avena fatua, 753 images), creeping thistle (Cirsium arvense, 118) and green foxtail (Setaria viridis, 99), captured with a DJI Mini 3 drone and a Google Pixel 5 smartphone across winter wheat at tillering to stem elongation stages. Images came from morning and midday light windows, supplemented with a public wheat-weed dataset to vary camera, height, background and soil conditions.

The team split the data 70/15/15 and scored models on unweighted per-class F1 rather than raw accuracy, arguing that a model which finds wild oat but misses creeping thistle has no value to a grower. They applied synthetic occlusion via random erasing at probabilities of 0, 0.25 and 0.5, erasing rectangles covering 2–40% of image area, while holding out naturally occluded field images for testing.

How did each architecture perform?

• ViT-B16 (transformer): test accuracy rose from 80% to 84% to 86% as occlusion increased; F1-scores ranged 0.80–0.96.
• PvTv2 (pyramid transformer): best baseline at roughly 82% with no occlusion, dipped to 74–75% at moderate occlusion, recovered to about 78%.
• ResNet-50 (CNN): stable but middling, holding 75–79% across occlusion levels with F1-scores near 62–69%.
• EfficientNet-B0 (CNN): worst performer, falling from about 66% to 53% to 46% as occlusion rose.

A multivariable regression found architecture type, not occlusion level, drove results: transformer models outperformed CNNs by an average 14.56 percentage points (p ≤ 0.001). For ViT-B16 specifically, occlusion significantly improved accuracy (β = 0.14, p ≤ 0.001, R² = 0.887), which the authors attribute to occlusion acting as regularisation during training. For EfficientNet-B0 the relationship ran the other way (β = −0.41, p ≤ 0.001).

What does this mean for input costs and deployment?

The compute profile tempers the headline. ViT-B16 required 16.74 GFLOPs per image — roughly twice EfficientNet-B0's 8.39 and five times ResNet-50's 3.32 — with mean inference of 17.06 seconds on the test hardware versus 1.82 seconds for PvTv2. The authors conclude ViT-B16 suits server-side or offline workflows, while sub-5-million-parameter models fit edge devices on field robots where moderate robustness is acceptable.

The stakes for spray accuracy are explicit. The researchers write that prediction instability "can compromise herbicide application accuracy, leading to either missed weed targets (false negatives) or blanket herbicide application to crop areas (false positives)," both of which undermine precision spraying economics and resistance management.

Caveats before growers bank on it

The authors flag real limitations. Dataset augmentation to boost sample size may have introduced bias and data-leakage risk; synthetic rectangular occlusion does not fully mimic irregular natural weed-crop overlap; and the study covered one season at one farm plus a public dataset. The design also cannot separate the contribution of self-attention from ViT-B16's larger compute budget, since the two co-vary.

They call for multi-season, multi-location datasets, temporal imagery sequences, and hybrid CNN-transformer architectures as the path to edge-deployable, real-time weed classification — with field trials on operational spraying platforms as the necessary next step before these gains translate into measurable herbicide savings.