ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

WeedExpert-R1: Incentivizing Domain-Specific Reasoning in MLLMs with Reinforcement Learning for Precision Weed Grounding

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2600412.(doi:10.13031/aim.202600412)
Authors:   Zonglin Yang, Wei-Zhen Liang, Nevin Lawrence, Xin Qiao, Chi-En Chiang, Fuchen Li
Keywords:   Precision agriculture, Weed grounding, Multimodal large language models, Reinforcement learning.

Abstract. Precision weed control requires both species-level identification and per-instance localization, but conventional object detectors are constrained by a closed-vocabulary paradigm that hinders cross-region deployment and lack the reasoning capability to justify their predictions in complex agricultural scenes. Recently, multimodal large language models (MLLMs) have demonstrated strong visual perception and reasoning in visual grounding, offering a promising alternative. However, their insufficient domain-specific botanical knowledge often leads to hallucinations during fine-grained weed identification. In this study, we introduce WeedExpert-R1, a multimodal reasoning model that, following the R1-style training paradigm, incentivizes visually grounded botanical reasoning through verifiable rewards. Specifically, a domain-specific, hallucination-free Chain-of-Thought (CoT) synthesis pipeline was proposed, which combines a human-curated botanical trait dictionary covering leaf shape, margin, petiole, and stem morphology, with an LLM workflow to automatically generate high-quality reasoning data for supervised fine-tuning as a cold start. Group Relative Policy Optimization (GRPO) was then applied with four verifiable rewards, including format, IoU, count, and length, to further enhance the model's capability to perceive and localize diverse weeds. WeedExpert-R1-4B was shown to outperform both frontier proprietary models (GPT-5.4, Gemini-3.1-Pro) and open-source baselines an order of magnitude larger (Qwen3-VL-30B-Instruct, Gemma-4-31B-it) by approximately +35 points across five metrics on 37 weed species. Moreover, because grounding accepts species names as runtime prompts rather than a fixed class list, WeedExpert-R1 demonstrated open-vocabulary capability on unseen species, supporting deployment across diverse regions and crops without retraining.

(Download PDF)    (Export to EndNotes)