Abstract
Effective weed management in vegetable production is increasingly challenged by labor shortages, rising production costs, and mounting pressure to reduce herbicide usage. Conventional deep learning models for crop-weed detection often suffer from performance degradation under the data-scarce scenarios commonly encountered in agricultural vision applications. Vision-language models (VLMs) offer a promising alternative by enabling text-guided detection that leverages semantic descriptions to improve crop-weed differentiation. This research evaluated the performance of three advanced VLM-based detectors, including Grounding DINO, YOLO-World, and Qwen3-VL, for crop-weed detection in vegetable production systems. To enrich textual representations automatically, Gemini 3 was employed to generate descriptive keywords for the training dataset, which were subsequently converted into text embeddings and integrated with image data during model training. Compared with YOLOv8-large, which achieved 91.0% mAP@50 and 72.2% mAP@50:95, text label-based Grounding DINO and YOLO-World demonstrated comparable in-season performance, achieving 91.3% mAP@50 and 73.1% mAP@50:95, and 91.0% mAP@50 and 72.5% mAP@50:95, respectively. In contrast, keyword-based prompts resulted in inferior performance for all three VLMs. Grounding DINO achieved the best cross-season generalization performance, with 60.9% and 45.6% mAP@50 on the 2021–2023 and 2024 datasets, respectively, improving mAP@50 by 12.1% and 7.7% over YOLOv8-large (48.8% and 37.9% mAP@50, respectively). These results suggest that the benefits of current large-scale VLMs are limited and conditional. While Grounding DINO provides advantages in cross-season generalization, YOLO-World and Qwen3-VL do not consistently improve performance, and enriched descriptions may even negatively affect model performance. The software programs developed in this study have been made publicly available to support future research and development in precision weed management.