Answers are grounded in this paper and cite their sources.
Precise phenotypic perception and reliable decision-making are critical for the intelligent management of greenhouse tomatoes. However, existing methods often lack an integrated closed-loop “perception-to-reasoning” workflow and struggle with high computational costs on edge devices. This study proposes PP-LLMs, an innovative framework that bridges visual perception and agronomic reasoning by integrating a lightweight detection model, YOLON, with a fine-tuned Large Language Model, Tomato-Qwen. Our core contributions are threefold. First, we constructed the TGMD (Tomato Growth Multimodal Dataset), a benchmark comprising 9,519 high-quality expert-annotated language pairs covering descriptions of tomato growth status and agricultural production knowledge.Second, we developed YOLON, which employs efficient DELK-Block and GSLK modules to achieve high-precision detection of tomato organs while maintaining a lightweight architecture suitable for real-time edge deployment. Third, by applying a two-stage fine-tuning process (LoRA and prompt learning) to Qwen-7B-Chat, we enabled Tomato-Qwen to perform complex agronomic reasoning based on real-time phenotypic and environmental data. Experimental results demonstrate that this strategy significantly outperforms mainstream baselines. Tomato-Qwen achieved substantial improvements in QA tasks, with BLEU-4 and ROUGE-L scores increasing by 29.42% and 33.13% respectively.This research demonstrates the preliminary effectiveness of visual-language collaboration in agriculture and suggests a potentially scalable solution for autonomous greenhouse management. But the generated agronomic suggestions should be regarded as preliminary candidate outputs whose practical safety and reliability have not yet been validated by domain experts.
No takes yet. Share an insight, caveat, or question.
Zhao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: