Scholarlink

Purpose <p>This study addresses the challenge of real-time detection of the weeds <i>Colchicum autumnale</i> and Rumex species on grassland sites, which is an inherently difficult problem because the predominantly green weed leaves provide little contrast to the similarly colored vegetation backgrounds. The resulting detector will be integrated into the SELBEWAG tool, a non-chemical, site-specific weed treatment device.</p> Methods <p>We collected and annotated RGB video recordings from grassland sites in Southwest Germany and trained a quantized EfficientDet object detection model, which has been optimized for low latency on edge devices.</p> Results <p>The detection system achieved a mean average precision of 0.606 across both weed types (0.617 for Rumex and 0.595 for <i> C. autumnale</i>). With an optimal decision threshold, the model demonstrated precision values of 56.0% for <i> C. autumnale</i> and 48.1% for Rumex, with corresponding recall values of 62.1% and 67.1%, respectively. Detection performance was influenced by surrounding vegetation height and weed clustering.</p> Conclusions <p>The developed system provides effective real-time detection of grassland weeds suitable for integration with the SELBEWAG tool. While detection challenges remain, in particular in high vegetation conditions, the approach significantly improves upon area-wide treatment methods by targeting only the necessary area rather than entire fields.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Real-time detection of Rumex and C. autumnale in grasslands

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月9日

查看全文
  • Lukas Petrich,
  • Ingo-Leonard Haußmann,
  • Georg Lohrmann,
  • Albert Stoll,
  • Volker Schmidt

摘要

Purpose

This study addresses the challenge of real-time detection of the weeds Colchicum autumnale and Rumex species on grassland sites, which is an inherently difficult problem because the predominantly green weed leaves provide little contrast to the similarly colored vegetation backgrounds. The resulting detector will be integrated into the SELBEWAG tool, a non-chemical, site-specific weed treatment device.

Methods

We collected and annotated RGB video recordings from grassland sites in Southwest Germany and trained a quantized EfficientDet object detection model, which has been optimized for low latency on edge devices.

Results

The detection system achieved a mean average precision of 0.606 across both weed types (0.617 for Rumex and 0.595 for C. autumnale). With an optimal decision threshold, the model demonstrated precision values of 56.0% for C. autumnale and 48.1% for Rumex, with corresponding recall values of 62.1% and 67.1%, respectively. Detection performance was influenced by surrounding vegetation height and weed clustering.

Conclusions

The developed system provides effective real-time detection of grassland weeds suitable for integration with the SELBEWAG tool. While detection challenges remain, in particular in high vegetation conditions, the approach significantly improves upon area-wide treatment methods by targeting only the necessary area rather than entire fields.