DINOv3 Meets YOLO26 for Weed Detection in Vegetable Crops

Boyang Deng; Yuzhen Lu

cs.CVFeb 25, 2026

DINOv3 Meets YOLO26 for Weed Detection in Vegetable Crops

Authors: Boyang Deng, Yuzhen Lu

Organizations: Department of Biosystems & Agricultural Engineering, Michigan State University East Lansing, MI 48824

Abstract

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.

Figures & tables

Figure 1: The pipeline of DINOv3 training and evaluation.

Figure 2: The training pipeline of the DINO-YOLO framework for weed detection.

Backbone Configuration	Crop (%)	Weed (%)	Average (%)
Official ViT-Small	92.2	83.1	87.7
+ Pretraining	95.0	83.7	89.3
+ Gram Anchoring	96.6	83.3	89.9
+ High-Res Adaptation	88.6	76.4	82.5

Table 1: Linear probing accuracy of the DINOv3-trained ViT-Small backbone. The finetuning is based on the DINOv3 -trained official ViT-small model, consisting of three consecutive phases: pretraining, gram anchoring, and high-res adaptation. Crop refers to the average accuracy across all crop classes, while weed refers to the mean accuracy for all weeds.

Components						Metrics
Y-H	Y-H*	Y-B	ViT	STA	Align	Precision (%)	Recall (%)	$\text{mAP}_{50}$ (%)	$\text{mAP}_{50:95}$ (%)	Lat. (ms)	Param. (M)
✓		✓				80.8	79.7	84.9	63.2	12.0	24.8
	✓	✓				86.3	84.3	89.8	68.9	12.4	24.5
	✓		✓			89.7	87.2	91.9	71.6	35.1	36.1
	✓		✓	✓		87.7	87.8	91.5	72.6	36.1	36.7
	✓	✓	✓			89.6	86.5	91.3	71.0	41.6	50.2

Table 2: Ablation study to evaluate the contributions of different components in the DINO-YOLO framework. Y-H refers to the YOLO26-large head, and Y-H* indicates the modified YOLO26-large head. Y-B denotes the original YOLO26-large backbone. ViT represents the DINOv3-pretrained ViT-small. STA is the auxiliary branch added to the ViT, and Align refers to the feature alignment loss used in the dual-branch backbones. Lat. stands for latency, and Param. denotes model parameters.

Dataset	Model	Precision (%)	Recall (%)	$\text{mAP}_{50}$ (%)	$\text{mAP}_{50:95}$ (%)
2025	YOLO26	81.8	81.0	$86.9^{b}$	$66.2^{b}$
	YOLO26*	86.8	84.1	$90.3^{a}$	$69.4^{ab}$
	DINO-YOLO26*	89.6	87.2	$\textbf{92.3}^{a}$	$\textbf{72.3}^{a}$
	DINO-YOLO26	88.6	87.7	$92.2^{a}$	$71.4^{a}$
2024	YOLO26	42.8	30.8	$29.6^{c}$	$14.6^{b}$
	YOLO26*	44.4	34.9	$33.9^{b}$	$16.2^{c}$

Table 3: Performance comparison of YOLO26 and DINOv3-YOLO26 models across different year datasets. Three replications are used. YOLO26* indicates the modified YOLO26 model trained with DFL loss, and DINO* denotes that the DINO-ViT model is fine-tuned on the weed dataset. Superscripts denote the Compact Letter Display (CLD) generated from Tukey’s HSD multiple comparison test, where groups sharing a letter are not significantly different.

Weed Detection

Weed detection uses computer vision to locate and identify weeds in crops for targeted monitoring and control.

See all 13 papers

A Multi-Dataset Benchmark of YOLO-Based Weed Detection in Precision Agriculture
Sep 27, 2026Hristina Zdraveska, Vlatko Spasev, Ivica Dimitrovski +2Multi-Source Domain Adaptation Weed Detection
Faculty of Computer Science and Engineering, University Ss Cyril and Methodius, Rudzer Boshkovikj 16, P.O. 393, Skopje, 1000, North Macedonia
Evaluation of Convolutional and Transformer-Based Detectors for Weed Detection in Tomato Plantations
Apr 29, 2026Alcides Toledo Espinosa, Gerardo Antonio Álvarez Hernández, Ángel Eduardo Zamora-Suárez +2Agricultural Image Analysis Weed Detection
Instituto Politécnico Nacional · CIDETEC-IPN · Mexico City, Mexico +1
Vision-Language Semantic Grounding for Multi-Domain Crop-Weed Segmentation
Feb 27, 2026Nazia Hossain, Xintong Jiang, Yu Tian +3Domain Generalization for Semantic Segmentation Semantic Segmentation
Department of Bioresource Engineering, McGill University, Sainte-Anne-de-Bellevue, QC, Canada · Department of Plant Science, McGill University, Sainte-Anne-de-Bellevue, QC, Canada

Sep 27, 2026J/K move · Enter open · S save

A Multi-Dataset Benchmark of YOLO-Based Weed Detection in Precision Agriculture

Hristina Zdraveska, Vlatko Spasev, Ivica Dimitrovski, Ivan Kitanovski+1

Faculty of Computer Science and Engineering, University Ss Cyril and Methodius, Rudzer Boshkovikj 16, P.O. 393, Skopje, 1000, North Macedonia

Multi-Source Domain Adaptation Weed Detection Cross-Domain Generalization Object Detection Precision Agriculture

Open paper AskPDF arXiv

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 mAP

_{50:95}

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.

Abstract

Weed Detection

Explore similar work

A Multi-Dataset Benchmark of YOLO-Based Weed Detection in Precision Agriculture

Evaluation of Convolutional and Transformer-Based Detectors for Weed Detection in Tomato Plantations

Vision-Language Semantic Grounding for Multi-Domain Crop-Weed Segmentation