Organizations: Department of Biosystems and Agricultural Engineering, Michigan State University, East Lansing, MI 48824, USA · Department of Horticulture, Michigan State University, East Lansing, MI 48824, USA
Machine vision-guided precision spraying is promising for sustainable vegetable production by enabling site-specific weed targeting, reducing herbicide application while protecting crop yield and quality. However, the effects of operating speed of sprayer systems on the integrated detection-tracking-actuation pipeline remain insufficiently characterized under field conditions. This study evaluated a semi-autonomous, tractor-based spraying system for targeted weed control in radish and quantified performance across three operating speeds. A five-season dataset comprising 10,599 images and 84,311 annotated instances across seven classes was curated for model development. A YOLOv11-medium model was integrated with ByteTrack tracking and a customized 12-nozzle sprayer on a semi-autonomous electric tractor. Field trials were conducted at 0.13, 0.27, and 0.45 m/s (i.e., 0.3, 0.6, and 1.0 mph) to characterize speed-dependent changes in weed detection, targeting, actuation, and crop damage. At the lowest speed of 0.13 m/s, the system achieved 94.6% mAP@50 and 93.2% recall for weed and crop detection. However, increasing platform speed substantially reduced system performance, with weed hit rate declining from 78.3% to 35.0% and crop contact rates increasing from 20.1% to 28.7%. Consequently, the system did not meet predefined commercial-viability thresholds. These results demonstrate that practical deployment requires improvements beyond weed detection alone, particularly coordinated improvements in motion-robust plant perception, spray-timing compensation, and crop-adjacent actuation precision. The identified performance bottlenecks provide specific engineering targets for advancing vision-guided precision spraying toward practical, high-speed vegetable weed management. The software programs developed in this study have been made publicly available to support future precision weeding research.
Figures & tables
Fig. 1: Semi-autonomous precision spraying platform: (a) photograph of the converted electric tractor (Oggún II–based; Section II-B ) and (b) schematic showing navigation, imaging with onboard computing, and the sprayer actuator. Total platform length: 5.06 m (16.6 ft), comprising a 3.81 m tractor unit and a 1.25 m spray extension.
Fig. 2: System architecture, organized into three functional subsystems. (1) Vision and sensing: an enclosed imaging chamber acquires frames at 30 frames per second, which are preprocessed and passed to the YOLOv11-medium detector with ByteTrack tracking. (2) Spray timing control: platform velocity is estimated from Hall Effect wheel pulses and pixel-based tracking, median filtered, and used to compute the spray delay (Eq. ( 8 )) and valve open duration (Eq. ( 10 )) for each nozzle. (3) Hardware integration: onboard computer, Arduino relay box, pump, and dual manifolds driving 12 independently actuated nozzles at three operating speeds (0.3–1.0 mph). The control-time pipeline enforces a ≤ 150 ms total response window.
Fig. 3: Nozzle–solenoid valve–manifold assembly used in this study: (a) schematic and (b), (c) photographs.
Dataset
Images
Instances
Species/classes
Collection period
Primary imaging platform
3SeasonWeedDet10 a
8,436
27,963
10 weed species
2021–2023
Smartphones, digital cameras, vehicle-mounted
Weed&CropDet2024
1,002
37,010
5 crops, 12 weeds
2024
Vehicle-mounted platform
Weed&CropDet2025
1,161
19,338
Radish, 6 weeds
June–August 2025
Tractor-mounted system (this study)
Combined total
10,599
84,311
—
2021–2025
Multiple platforms
TABLE I: Dataset Composition Across Five Growing Seasons
Class name
Scientific name
Role
Notes
Radish
Raphanus sativus L.
Crop
Primary crop requiring herbicide protection
Carpetweed
Mollugo verticillata L.
Weed
Summer annual broadleaf
Lambsquarters
Chenopodium album L.
Weed
Summer annual broadleaf
Palmer amaranth
Amaranthus palmeri S. Watson
Weed
Summer annual broadleaf
Smartweed
Persicaria spp.
Weed
Summer annual broadleaf
Purslane
Portulaca oleracea L.
Weed
Summer annual broadleaf
TABLE II: Target Species and Classes in the Combined Dataset
Fig. 4: Overview of the YOLOv11 training pipeline and inference architecture. Three model variants (nano, small, and medium) were compared on the five-season dataset; YOLOv11-medium was selected for field deployment based on its superior mAP@50 (89.9%) and real-time compatibility at 30 FPS. At inference, the model applies post-processing filters, removes duplicate detections, and tracks objects with ByteTrack.
Fig. 5: Indoor validation of the precision spraying system. (a) Real-time software interface [ 15 ] showing the YOLOv11 detection output, the status of all 12 valves, and processing throughput (FPS). (b) Sand bin test apparatus (0.6 m × 2.4 m) with 100 white polyester targets (5 cm × 5 cm) at approximately 5 targets/m 2 ; blue dye deposition after each pass was used to quantify weed hit rate and spray positioning accuracy.
Fig. 6: Field trials conducted in August 2025. (a) The semi-autonomous spraying platform traversing radish rows at the MSU research farm. (b) Real-time software interface showing YOLOv11 detections of weeds (green boxes) and radish (blue boxes). (c)–(e) Close-up of inter-row weed populations targeted during the trials; blue dye deposition on soil and plant surfaces after a sprayer pass; and radish canopy at the 5–7 true-leaf stage.
Metric
Definition
Formula
Weed hit rate (WHR)
Percentage of ground-truth weeds receiving spray contact ( IoU>0 ).
WHR=NwgtNwhit×100%
Crop contact rate (CCR)
Percentage of ground-truth crops receiving any spray contact.
CCR=NcgtNchit×100%
Weed precision rate (WPR)
Percentage of total spray area deposited on weed targets.
WPR=ATAw×100%
Crop precision rate (CPR)
Percentage of total spray area deposited on crop plants.
CPR=ATAc×100%
Soil precision rate (SPR)
Percentage of total spray area deposited on bare soil.
SPR=ATAb×100%
Average aligned distance ( davg )
Mean Euclidean distance between the spray centroid and the weed centroid for successfully hit weeds (mm).
davg=N1i=1∑NSCi−WCi
TABLE III: Spray Targeting and Crop Protection Metrics Used for System Evaluation
Criterion
Threshold
Rationale
Source
Weed control efficacy
> 90%
Conventional herbicides achieve 90–99% control; precision systems must approach this standard to minimize supplemental manual weeding.
Slaughter et al. [ 36 ]
Crop damage
< 10%
Economically acceptable yield-loss threshold for vegetable production with narrow profit margins.
Slaughter et al. [ 36 ]
Operating speed
≥0.67 m/s (1.5 mph)
Minimum for commercial field efficiency in specialty-crop operations; enables timely weed control within critical windows.
Vijayakumar et al. [ 12 ]
TABLE IV: Commercial Viability Thresholds for Precision Spraying Systems
Model
Inference time (ms)
Crops mAP@50 (%)
Weeds mAP@50 (%)
YOLOv11-nano
10.0
89.8
82.5
YOLOv11-small
14.6
87.1
87.2
YOLOv11-medium
20.9
90.7
89.1
TABLE V: YOLOv11 Variant Performance on the Five-Season Dataset
Test condition
Precision (%)
Recall (%)
mAP@50 (%)
Indoor (0.27 m/s)
98.1
99.5
99.1
Field (0.13 m/s)
95.8
93.2
94.6
Field (0.27 m/s)
88.4
80.7
84.9
Field (0.45 m/s)
76.3
62.5
70.1
TABLE VI: Detection Performance Across Operating Speeds
Fig. 7: Speed–accuracy trade-offs across the three operating speeds in radish production: weed hit rate (WHR), crop contact rate (CCR), and detection accuracy (mAP@50). Dashed lines mark the commercial-viability thresholds (WHR > 90%, CCR < 10%); vertical arrows give the shortfall of WHR relative to the 90% target in percentage points.
Indoor
Field
Metric
0.27 m/s
0.13 m/s
0.27 m/s
0.45 m/s
Weed hit rate (%)
82.0
78.3
58.7
35.0
Crop contact rate (%)
—
20.1
24.3
28.7
Translation efficiency (%)
82.4
84.0
72.7
56.0
Selectivity ratio
—
3.9:1
2.4:1
1.2:1
Weed precision rate (%)
77.0
90.1
82.4
71.6
TABLE VII: Spraying Performance Across Operating Speeds
Developing robust models for precision vegetable weeding is currently constrained by the scarcity of large-scale, annotated weed-crop datasets. To address this limitation, this study proposes a foundational crop-weed detection model by integrating heterogeneous datasets and leveraging self-supervised learning. A total of 618,642 crop-weed images were initially collected and subsequently refined to 199,388 filtered images for fine-tuning a DINOv3 vision transformer (ViT-small) through a sequential curation strategy. The fine-tuned DINOv3 backbone was then integrated into YOLO26, serving either as a primary backbone or part of a dual-backbone architecture. A feature alignment loss was introduced in the dual backbone framework to enhance feature fusion with minimal computational overhead. Experimental results show that the proposed DINOv3-finetuned ViT-small-based YOLO26-large achieved up to a +5.4% mAP50 gain on in-domain images collected in the 2025 season. Moreover, it demonstrated strong cross-domain generalization with mAP50 improvements of +14.0% on the 2021-2023 season dataset and +11.9% on the 2024 season dataset, compared to the standard YOLO26-large. Although the DINOv3-YOLO26-large model has 45.6% more parameters and a 2.9x increase in inference latency, it maintains real-time performance at ~28.5 frames per second (fps). The curated dataset and software programs developed in this study will be made publicly available.
Boyang Deng, Yuzhen Lu
Department of Biosystems & Agricultural Engineering, Michigan State University East Lansing, MI 48824
Weed detection is an important component of precision agriculture, enabling site-specific weed management and reducing unnecessary herbicide use. Although deep learning methods have achieved strong results for crop and weed detection, many studies rely on single-dataset evaluation, making it difficult to assess robustness across different agricultural domains. This paper presents a multi-dataset benchmark of deep object detectors for weed detection in precision agriculture, with a focused evaluation of YOLO26 models. We evaluate nano, small, and medium variants on seven public weed-detection datasets covering different crops, weed species, field conditions, acquisition setups, and annotation protocols. The models are compared in terms of detection accuracy, model complexity, inference latency, FPS, and model size. In addition to in-dataset evaluation, we investigate cross-domain generalization using a unified one-class weed setup and evaluate multi-source training using the combined training subsets from all datasets. The results show that YOLO26 achieves strong in-dataset performance, with YOLO26m obtaining the highest average accuracy and YOLO26s providing the best practical accuracy-efficiency trade-off. However, cross-domain performance decreases substantially, with YOLO26s dropping from an average in-domain mAP50:95 of 0.603 to 0.148 in the off-domain setting. Multi-source training improves performance on several datasets, but does not fully eliminate domain shift. Overall, the benchmark highlights the importance of dataset diversity, domain similarity, and target-domain adaptation for robust weed detection in real-world precision agriculture applications.
This paper presents a comparative evaluation of convolutional and transformer-based object detection architectures for early weed detection in tomato plantations. Representative models from each paradigm are considered, including YOLOv26-nano, a recent variant of the YOLO family, and RT-DETR Large and RF-DETR Medium as transformer-based architectures. The evaluation was conducted on the GROUNDBASED_WEED dataset, considering six weed classes and an additional category corresponding to unidentified plants, which allowed for the assessment of performance in terms of detection accuracy and computational efficiency using metrics such as precision, recall, average precision, and inference speed, as well as non-parametric statistical tests. The results highlight a clear trade-off between efficiency and contextual modeling: CNN-based detectors achieve high performance at a lower computational cost, while transformer-based approaches offer better global context capture at the expense of higher resource demands. These results provide practical criteria for model selection in precision agriculture applications.