Large language model assisted decision support framework for uncertainty aware detection and management of tomato lateral shoots.

Abstract

Accurate identification of tomato lateral shoots is essential for automated pruning and plant monitoring in greenhouse production. However, complex illumination, leaf occlusion, and morphological variability often reduce detection reliability in optical vision systems. This study proposes an optical vision-based framework that integrates deep learning perception with large language model assisted pruning decision support. A tomato lateral Shoot image dataset was constructed using RGB imaging in greenhouse environments. A lightweight YOLOv8n instance segmentation model with the Convolutional Block Attention Module (CBAM) was developed to enhance feature representation. Data augmentation strategies were applied to simulate illumination variations and improve model robustness. Model interpretability was analyzed using Principal Component Analysis (PCA) and Gradient weighted Class Activation Mapping (Grad CAM). Experimental results show that the proposed YOLOv8n-seg+CBAM model achieves a mAP 0.5 of 98.1% with only 3.28M parameters and an average inference time of 8.0 ms per image. Monte Carlo Dropout was further introduced to estimate the spatial uncertainty of cutting points. These structured perception features were provided to a large language model (LLM), enabling context aware pruning decision assistance. The proposed framework integrates vision-based shoot detection, uncertainty estimation, and LLM-assisted reasoning into a unified pipeline, enabling more reliable pruning decisions and improving safety and robustness compared with vision-only approaches in greenhouse environments.

Plant phenotyping relevance

トマト側枝をRGB画像から検出・セグメンテーションし、不確実性推定まで行う画像ベースの植物形態計測手法を開発しており、方法論が中心である。

Code and data availability

The paper describes a custom tomato lateral shoot image dataset (1,130 annotated RGB greenhouse images) and a YOLOv8n-seg+CBAM analysis pipeline, but no public repository, deposit, or authors' URL for the dataset, code, or trained model is provided. The data availability statement only promises raw data from theauthors

No evidence-backed public reproduction asset is currently recorded.